Lake Piburg covers about 0.13 square kilometres, reaches a maximum depth of roughly 24.6 metres and sits at 913 metres above sea level. The lake undergoes pronounced seasonal thermal stratification, with a relatively stable hypolimnion during the stratified period and complete or near-complete overturn during spring and autumn mixing. Its high-frequency dataset came from in situ dissolved-oxygen and temperature loggers deployed at depths from 1 to 24 metres. The lower-frequency dataset consisted of samples collected at 3-metre intervals and measured by Winkler titration. For comparison, the researchers interpolated both datasets onto a common 1-metre depth grid and calculated daily layer-level values.

When the approximately monthly profiles were interpolated through time, their volume-weighted oxygen concentrations followed the major seasonal trajectory in all three operational depth zones. Pearson correlations between the reconstructed monthly series and the hourly-sensor series were 0.928 in the epilimnion, 0.945 in the metalimnion and 0.916 in the hypolimnion. Those strong correlations did not mean the measurements were identical. Root mean square errors were 0.67, 0.84 and 1.09 milligrams per litre, respectively, while mean differences between sensor and reconstructed values were +0.25 milligrams per litre in the upper layer, −0.57 in the middle layer and −0.97 in the deep layer. The hypolimnion therefore showed both the lowest correlation and the largest negative bias among the three layers.

Differences were particularly pronounced at individual depths close to the boundary between the metalimnion and hypolimnion. In an analysis of 229 paired date-depth observations from 29 original sampling dates, the largest root mean square error occurred at 9 metres. The mean sensor-minus-profile difference there was −2.49 milligrams per litre, and the maximum absolute difference reached 6.61 milligrams per litre on 5 November 2024. At 12 metres, the mean difference was −1.42 milligrams per litre. Errors were much smaller at most other depths, including 0.23 milligrams per litre at 24 metres. The study’s authors said that differences in sampling depth, sensor and sampler positions, vertical integration, moving gradients and analytical uncertainty could all contribute, but the available data could not separate those effects quantitatively.

The clearest divergence appeared in estimates of hypolimnetic oxygen depletion rate. The rate was calculated by regressing daily, volume-weighted deep-water oxygen concentration against elapsed time during sustained depletion periods. The full hourly-sensor reference rates were −0.0170, −0.0187 and −0.0319 milligrams per litre per day in 2022, 2023 and 2024, respectively. To isolate the effect of temporal sampling, the researchers repeatedly thinned the hourly record to 7-, 14-, 30- and 60-day intervals, shifting the first sampling date through every possible starting-day offset within fixed annual fitting periods. These schedules produced both faster and slower depletion estimates than the full hourly record, demonstrating that the inferred slope depended on where the sparse observations happened to fall within the depletion period.

The average within-year range of signed deviations from the full-sensor reference expanded sharply as sampling intervals lengthened. It was 8.0 percentage points for 7-day sampling, 20.5 percentage points for 14-day sampling, 47.5 percentage points for 30-day sampling and 48.6 percentage points for 60-day sampling. The increase was not strictly monotonic in every year, and the 60-day result was based on fewer estimable schedules because several long-interval fits relied on only two sampling dates. Traditional regression goodness-of-fit measures could not be calculated for those schedules, so the researchers used deviation from the full hourly-sensor depletion rate as the primary indicator of sensitivity. A separate comparison using the actual monthly sampling dates introduced both frequency and method differences, because the monthly reconstruction relied on discrete Winkler measurements rather than in situ sensors.

Depth-specific counts of days with oxygen below 2 milligrams per litre showed a more robust vertical pattern than the absolute duration of low-oxygen conditions. Counts generally increased with depth, and no daily mean below the threshold appeared at depths shallower than 10 metres in either dataset. The shallowest affected depth in the sensor record was 10 metres in 2022, 11 metres in 2023 and 10 metres in 2024, compared with 12 metres in every year for the reconstructed monthly profiles. Correlations between paired depth-wise counts were 0.947, 0.933 and 0.977 for 2022, 2023 and 2024, respectively, using 24 paired depths per year. Both approaches therefore identified the depths where prolonged oxygen deficits occurred, even when the estimated number of affected days differed.

Those threshold counts cannot be treated as equivalent full-year exposure estimates. The sensor dataset contained 212, 317 and 324 available days in 2022, 2023 and 2024, respectively, while the interpolated monthly record covered 263, 365 and 340 days. Root mean square differences between depth-wise counts were 40.1, 55.9 and 29.7 days, and mean reconstructed-minus-sensor differences were +5.6, −15.0 and −12.3 days. Because the observation windows and missing-data patterns differed, the analysis compared the shape of oxygen exposure through the water column rather than providing an unbiased estimate of annual low-oxygen duration. The authors used the counts descriptively because adjacent depths in a reconstructed profile are not independent replicates.

The findings support metric-specific monitoring rather than a universal optimum sampling interval. Monthly profiles remained informative for seasonal context, historical comparison and locating persistent low-oxygen layers, and they retained much of the broad lake-wide oxygen trajectory. Higher-frequency observations became more valuable when investigators wanted to estimate depletion rates, identify the beginning or end of depletion, detect reoxygenation or characterize short-lived changes between visits. The study did not establish whether monthly observations alone would identify the same depletion windows as the hourly record. Instead, the researchers selected fitting periods from the high-frequency data and held them fixed during subsampling, so the analysis isolated sensitivity within a known depletion interval rather than testing the detection of that interval.

Generalization beyond Lake Piburg remains uncertain. The study examined one small lake over three years, whereas oxygen dynamics can differ with depth, trophic state, mixing regime, climate and the timing of stratification. The authors therefore described their results as evidence that sampling frequency should be matched to the metric being estimated, not as a prescription for all lakes. For Lake Piburg specifically, the study indicates that monthly profiling can remain valuable for broad seasonal and long-term assessment, while denser observations provide more stable estimates of deep-water oxygen loss and better resolve rapid depletion and replenishment events. The findings also preserve the value of historical monthly records, which were collected before continuous sensors became practical for routine lake monitoring.

Subject of Research: Effects of monitoring frequency on dissolved-oxygen metrics in a seasonally stratified lake

Article Title: Robustness of oxygen metrics to sampling frequency in a stratified lake

Article References: Weniger, E., & Sommaruga, R. (2026). Robustness of oxygen metrics to sampling frequency in a stratified lake. Environmental Monitoring and Assessment, 198(11), Article 1239. https://doi.org/10.1007/s10661-026-16081-8

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16081-8

Keywords: lake monitoring, dissolved oxygen, sampling frequency, oxygen depletion, high-frequency sensors, depth profiles, Lake Piburg, environmental sensors, water quality, seasonal stratification, Robustness, oxygen

The researchers describe the work as evidence that frequency-domain information can complement conventional spatial features, although their analysis identifies important limitations in cross-dataset generalization and baseline comparisons. Modern image generators and editing tools can produce forgeries whose visible imperfections are subtle or absent, making it increasingly difficult to distinguish synthetic material from authentic photographs. Existing detectors often concentrate on spatial cues such as inconsistent textures, illumination, edges and object boundaries.

Those signals can be effective, but they may not transfer well when a detector encounters a generator or manipulation process that it did not encounter during training. The new study focuses on frequency-domain anomalies, patterns that can be difficult to see in the image itself but become accessible after converting feature maps into their spectral representation. MSFEN is built on a pretrained ConvNeXt-Tiny encoder with four stages.

A Fourier Enhancement Module is placed after each stage, giving the network access to both spatial and frequency-domain information at several levels of abstraction. For an input feature map, the module applies a real fast Fourier transform, separates the real and imaginary components of the resulting spectrum, and concatenates them along the channel dimension. A learnable 1-by-1 grouped convolution then modulates the two components independently.

The transformed representation is reconstructed as a complex signal and returned to the spatial domain with an inverse real fast Fourier transform, preserving the original height and width. The module does not stop at the inverse transform. Its output passes through depthwise convolution, a GELU activation and a channel-mixing sequence that expands the number of channels fourfold before compressing them again.

A residual connection combines this result with the original feature map. In this way, the module can amplify potentially informative spectral patterns while retaining the local spatial information carried by the original representation. The design treats frequency and space as complementary rather than competing views of an image: high-frequency components are associated with fine detail and texture, while lower-frequency components provide broader structural information.

Because manipulated regions can range from a small edited object to an entirely synthetic image, MSFEN combines representations from all four ConvNeXt stages. Its Multi-Scale Feature Fusion Module uses 1-by-1 convolutions to align channel dimensions, upsamples the feature maps to a common resolution and concatenates them. Two 3-by-3 convolutional layers then produce a unified representation containing both local texture evidence and higher-level semantic context.

An analysis using one to four backbone stages found that average detection accuracy improved as more scales were incorporated, although the gains began to saturate at four stages. The authors suggest that a three-stage configuration could offer a better balance between accuracy and computational cost in resource-constrained deployments. The fused features feed two related outputs.

One head classifies the entire image as authentic or forged, combining pooled statistical features with a flattened spatial representation. A second head generates a pixel-level forgery mask by convolving and upsampling the fused map to the input resolution. The final training objective combines binary cross-entropy losses for image classification, pixel localization and auxiliary image-level predictions attached to each of the four encoder stages.

These learnable loss weights all began at 1.0 and converged after 100,000 training steps to a ratio of approximately 1:1.07:0.92, indicating that the three objectives contributed at similar levels during optimization. For training and evaluation, the researchers used more than 5,000 forged images from CASIAv2, which includes splicing and copy-move manipulations, and generated 10,000 synthetic images each with ADM and BigGAN. The test sets consisted of the edited-image datasets CASIAv1, Columbia and NIST16, together with 3,000 generated images from ADM, BigGAN and VQDM.

CASIAv2 served as the image-editing dataset used in the experimental setup, whereas the reported editing-benchmark results came from CASIAv1, Columbia and NIST16. The model was trained for 100 epochs on two NVIDIA RTX 3090 graphics processors, using 512-by-512 inputs, AdamW, cosine annealing and a batch size of eight. Baseline models were retrained with the same data split, preprocessing, augmentation and optimization protocol as MSFEN.

Image-level and pixel-level outputs were thresholded at 0.5. The comparison included F3Net, MVSS, PSCC, TruFor, Mesorch and NFA-ViT. The experimental design therefore measured performance under a common training and evaluation setup, although the methods retained their original architectural differences.

The study also supplied pixel-level supervision: public editing datasets used their provided binary masks, fully synthetic images received masks containing all ones, and authentic images received masks containing all zeros. On the unseen VQDM dataset, MSFEN achieved an F1 score of 0.571, accuracy of 0.625 and area under the receiver operating characteristic curve of 0.848. The next-best reported AUC was 0.641 for NFA-ViT, illustrating that performance fell sharply for every tested method when the generative source changed.

The authors characterize the result as preliminary evidence of cross-generator generalization rather than proof that the system is robust to all new generators. On the editing benchmarks, MSFEN achieved the best reported accuracy on NIST16, at 0.634, while Mesorch performed slightly better on CASIAv1. The relatively low absolute performance on NIST16 shows that regional editing remained more difficult than recognition of fully synthetic images.

Stress tests applied Gaussian blur, Gaussian noise and JPEG compression to ADM-generated images. The network reported average accuracies of 0.967 under Gaussian blur, 0.993 under Gaussian noise and 0.993 under JPEG compression, the highest averages among the compared systems for each perturbation. Under particularly strong blur with a kernel size of 23, its accuracy remained 0.914.

The authors caution that these comparatively stable results partly reflect the nature of the test set: generative traces are distributed throughout the image rather than concentrated in a small edited region. At 512-by-512 resolution with a batch size of one, MSFEN contained 46.9 million parameters, required 121.7 billion floating-point operations and ran at 33 frames per second. It used about 45% fewer parameters than Mesorch at comparable FLOPs.

Several qualifications limit how broadly the findings can be interpreted. The baseline systems originated with different backbone architectures, and the proposed Fourier and fusion modules were integrated into a strong ConvNeXt backbone, so the observed gains cannot be completely separated from the benefits of that architecture. Failure cases on NIST16 showed that MSFEN can suppress background responses more effectively than several baselines, yet it still misses highly subtle edits or mistakes natural high-contrast textures for manipulation.

The authors identify broader cross-domain testing, emerging diffusion models and temporal analysis for video as future directions rather than demonstrated capabilities of the present system. Overall, the results support the approach as a promising computational trade-off, while leaving open how it will perform against future generators and more difficult forms of localized manipulation.

Subject of Research: A multi-scale Fourier neural network for detecting and localizing edited and AI-generated image forgeries

Article Title: A lightweight multi-scale fourier enhancement network for image forgery detection

Article References: Hou, Y., Miao, Z., Wang, J., Lu, X., Zhang, R., & Li, Y. (2026). A lightweight multi-scale fourier enhancement network for image forgery detection. Cybersecurity, 9(1), Article 244. https://doi.org/10.1186/s42400-026-00681-5

Image Credits: AI Generated

DOI: 10.1186/s42400-026-00681-5

Keywords: image forgery detection, Fourier analysis, multi-scale learning, ConvNeXt, generative AI, digital forensics, synthetic images, robustness, network efficiency, localization, lightweight, multi-scale

The technique, described in Scientific Reports, targets Wi-Fi fingerprinting, the dominant approach to indoor localization. Fingerprinting works by treating the pattern of received signal strength indicator (RSSI) values from all detectable access points as a unique signature of a physical location, much like a radio shadow cast by the surrounding infrastructure. A device measures which routers it can hear and how strongly, then matches that pattern against a pre-built database to estimate its position. The method requires no additional hardware and runs on commodity phones, which is precisely why it remains the most deployment-ready positioning technology available. But it carries a hidden cost: modern environments expose tens to hundreds of access points, producing high-dimensional, sparse fingerprints riddled with missing values that burden mobile processors, slow down scanning, and can degrade generalization when redundant or unstable signals creep into the training data.

The Egyptian team, led by Ahmed Y. Bahy El-Din, attacked this dimensionality problem with Shannon entropy, the foundational measure of information content introduced by Claude Shannon in 1948. For each access point, the researchers computed the entropy of its empirical RSSI distribution, discretizing observed signal values into bins and calculating how spread out the probability mass is. The logic is elegant: a router whose signal strength varies richly across different physical locations carries discriminative spatial information, while a router broadcasting a near-constant profile everywhere yields almost no positioning cues and merely injects noise into the fingerprint matrix. High-entropy access points earn their place in the feature set; low-entropy ones get pruned.

What distinguishes this work from earlier entropy-based filtering is its adaptivity. Previous schemes applied static percentile thresholds or fixed global cutoffs that ignored the reality that signal statistics differ dramatically from floor to floor. A conservative threshold preserves accuracy on sparse floors but fails to compress well-conditioned ones, leaving redundancy untouched. An aggressive threshold compresses stable floors but catastrophically discards vital spatial cues elsewhere. The researchers solved this by linking the selection percentile to a global coefficient of variation computed over the non-missing RSSI values on each floor, allowing the threshold to shift automatically according to measured signal dispersion. The scaling factor governing this shift was chosen through a minimum-sufficient-shift criterion, requiring that worst-case accuracy degradation across all floors stay below one percent while guaranteeing meaningful compression everywhere. Crucially, all entropy estimates and thresholds were computed exclusively on training folds during cross-validation, preventing any information leakage from held-out test samples.

The evaluation used the JUIndoorLoc dataset, a corpus of Wi-Fi fingerprints collected over 31 days from four Android devices across a three-floor building, spanning 172 distinct access points and multiple contextual conditions including open versus closed rooms and human presence versus absence. The team trained Random Forest classifiers with stratified five-fold cross-validation and swept static percentile thresholds from 10 to 95 percent to map the accuracy-dimensionality trade-off before engaging the adaptive rule. The results are striking. On Floor F3, the framework selected just 17 access points, a 90.11 percent reduction, with an accuracy change of only minus 0.41 percent. On Floor F4, it retained a mere 8 access points, compressing the feature space by 95.34 percent while losing just 0.51 percent accuracy. On Floor F5, 26 access points sufficed, an 84.88 percent reduction with a 0.42 percent accuracy loss. In every case, the adaptive threshold landed in a narrow operating region that fixed thresholds missed entirely.

The sensitivity analysis revealed why the adaptive approach matters. Floor F4, whose entropy distribution was highly concentrated, exhibited a sharp cliff effect: increasing the scaling factor from 5 to 40 collapsed accuracy from 0.9697 to 0.4378, a catastrophic 55 percent drop, because even small threshold shifts eliminated critical routers in that compact signal environment. Floors F3 and F5 degraded gradually by contrast. The chosen scaling factor of 5 proved Pareto-optimal, achieving the highest worst-case accuracy across all floors while delivering non-trivial compression on every one. Comparisons against standard entropy baselines with fixed 85 and 95 percent cutoffs, and against an Otsu-based adaptive threshold, confirmed the advantage: the conservative static threshold selected all 172 access points on F3 with zero compression, while the aggressive one caused a 15.35 percent accuracy collapse on that floor. The proposed method outperformed PCA-based dimensionality reduction on all three floors despite using dramatically fewer features, and unlike PCA, it preserved direct one-to-one correspondence with physical routers, enabling network administrators to identify which specific devices are critical for localization.

An ablation study addressed a subtle concern: Random Forest performs its own internal feature selection during node splitting, so was the observed benefit merely the classifier doing the pruning work? The researchers repeated the experiments with K-nearest neighbors, a classifier with no intrinsic feature-selection mechanism, and found that entropy-selected subsets preserved most performance despite reducing features from 172 to 17, 8, and 26 access points, with accuracy drops limited to 2.5, 0.5, and 0.4 percentage points respectively. Comparing entropy ranking against Random Forest’s own Gini importance ranking showed comparable or superior subset effectiveness across floors, establishing Shannon entropy as a classifier-agnostic, lightweight alternative to model-dependent importance measures.

Robustness testing under realistic degradation revealed a nuanced, floor-dependent trade-off. Under synthetic Gaussian RSSI noise, the entropy-selected subsets frequently equaled or surpassed the full feature space on Floors F3 and F5. On Floor F5, at noise levels of 1, 2, and 3 decibels, the selected subset achieved accuracies of 0.8472, 0.7748, and 0.7070, versus 0.7674, 0.7084, and 0.6659 for using all access points, demonstrating that pruning unstable signals actively filters noise. On Floor F3, the advantage held at mild noise but reversed under severe perturbation, where the redundancy of the full set acted as a buffer. Floor F4 told the opposite story: the extremely compact 8-access-point subset degraded far more rapidly under noise than the full set, whose redundancy provided vital resilience. Under random access point dropouts simulating infrastructure outages, however, the selected subsets proved exceptionally resilient everywhere, with Floor F3’s subset outperforming the full set by 3.2 percentage points at 30 percent dropout and Floor F4’s compact subset losing only 0.5 percentage points even under extreme feature compression.

The practical implications extend beyond accuracy. Benchmarking showed that pruning the feature space shrank model storage from 3.82 megabytes to between 0.24 and 0.73 megabytes, small enough for edge microcontrollers. Online inference latency on a simulated commodity smartphone dropped from 1.45 milliseconds to under 0.27 milliseconds per sample. Most remarkably, because the method isolates a compact subset of routers, mobile devices can execute targeted scans on only the channels where selected access points broadcast, cutting radio active time from roughly 300 milliseconds to about 60 milliseconds and yielding an estimated 80 percent reduction in scanning power consumption, a direct extension of battery life during active navigation.

The study also exposed honest limitations. Zero-shot transfer across environmental contexts failed almost completely, with near-zero accuracy across the vast majority of context pairs regardless of whether the full or selected feature set was used, revealing a fundamental distributional shift in RSSI under changing room configuration and occupancy rather than an artifact of feature selection. Yet a simple few-shot calibration protocol using just five target-domain samples per class restored accuracy to between 0.76 and 0.99 across all tested pairs, confirming that the selected subsets preserve the essential spatial structure needed for domain transfer. Device-specific entropy rankings showed high correlation with the pooled ranking, with mean Spearman correlation of 0.784 on the best-covered floor, suggesting that the most informative access points are identified consistently regardless of the measuring device. The authors caution that their evaluation rests on a single dataset from one building and that the optimal scaling factor may require recalibration elsewhere. Still, the core message stands: high-dimensional radio fingerprints are not always necessary for competitive indoor localization, and adaptive, information-theoretic filtering offers a scalable pathway toward deployable Wi-Fi positioning systems that are lighter, faster, and more robust than the brute-force alternative.

Subject of Research: Engineering

Article Title: Adaptive entropy-guided access point selection for robust multi-floor Wi-Fi fingerprinting

Article References: Bahy El-Din, A. Y., Nafea, H. B., Samra, A. S., & Abozied, M. A. H. (2026). Adaptive entropy-guided access point selection for robust multi-floor Wi-Fi fingerprinting. Scientific Reports, 16(1), 32706. https://doi.org/10.1038/s41598-026-72185-8

Image Credits: AI Generated

DOI: 10.1038/s41598-026-72185-8

Keywords: Wi-Fi fingerprinting, indoor localization, Shannon entropy, access point selection, adaptive thresholding, robustness, domain shift, Scientific Reports, Adaptive, entropy-guided, access, point

Digital signatures are the trust anchors of the modern internet. They prove that a message, a transaction, or a firmware update really came from the claimed sender and was not altered in transit. Nearly every signature scheme deployed today relies on mathematical problems that quantum computers running Shor’s algorithm could solve in polynomial time, which means the entire trust framework safeguarding blockchain networks, IoT ecosystems, and decentralized finance could collapse once sufficiently powerful quantum hardware arrives. The threat is compounded by harvest-now-decrypt-later strategies, in which adversaries record encrypted traffic today and decrypt it tomorrow with quantum machines. Signatures face a subtler version of the same problem: an attacker who records a public key now could forge signatures later, invalidating past transactions on an immutable ledger.

FSDS-OTS+ is founded on the computational difficulty of two lattice problems: Learning With Errors (LWE) and Short Integer Solution (SIS). These problems are believed to resist quantum attacks because no known quantum algorithm solves them efficiently, even in principle. LWE asks an attacker to distinguish noisy linear equations from random ones, while SIS requires finding a short vector that satisfies many random modular constraints. Both problems underpin worst-case to average-case reductions in lattice cryptography, meaning that breaking the scheme in the average case would imply solving the problems in their hardest known instances. The scheme integrates AI-enhanced public key compression, zero-knowledge verifiability, and homomorphic consistency checks to provide a compact and efficient post-quantum signature solution that goes beyond raw lattice hardness.

The methodology employs a dual approach. On the theoretical side, the authors provide reduction proofs that demonstrate EUF-CMA security against quantum attackers. EUF-CMA, or existential unforgeability under chosen-message attack, is the gold standard for signature security: it guarantees that no efficient adversary, even one allowed to request signatures on messages of its choosing, can produce a valid signature on any new message. Proving this in the QROM is significantly harder than in the classical random oracle model because a quantum adversary can query the oracle in superposition, forcing the reduction to handle quantum-accessible hash functions. The authors also demonstrate reduction robustness, ensuring that the security proofs remain valid even when the underlying assumptions are perturbed or the reduction parameters shift, a property that guards against subtle proof failures that have undermined other schemes.

On the empirical side, the team ran simulation-based assessments using MATLAB, Python with NumPy and PyTorch, and SageMath, incorporating real physical datasets to calibrate adversarial modeling. These datasets are unusual for a cryptography paper: GPS satellite time-drift offsets ranging from 10^-9 to 10^-6 seconds from the US Naval Observatory, muon decay lifetimes from CERN’s muon storage ring with a Lorentz factor of 29, and discrepancies from NIST optical lattice clocks accurate to 10^-18 seconds. The authors use these empirical phenomena of time instability to model quantum-adaptive drift, in which an attacker exploits tiny timing discrepancies to erode the nominal security strength of the scheme. By anchoring adversarial simulations in measured physical data rather than synthetic noise, the team aims to capture realistic attack surfaces that purely theoretical analyses might miss.

The experimental findings are striking. FSDS-OTS+ maintains at least 90 percent of its nominal security strength under quantum-adaptive drift settings, a robustness margin that few competing schemes have demonstrated under equivalent conditions. The scheme produces compressed public keys as small as 4.2 kilobytes, signature sizes of 6.8 kilobytes, and verification latency as low as 1.8 milliseconds. These numbers matter because post-quantum signatures are notoriously bulky: lattice and hash-based schemes often produce signatures tens of kilobytes in size, which strains bandwidth-limited IoT networks and inflates blockchain storage requirements. The AI-enhanced compression pipeline appears to be the key differentiator, learning to represent public keys more efficiently without sacrificing the hardness guarantees of the underlying lattice problems.

Comparative benchmarks indicate enhanced performance compared to SPHINCS+, Picnic, and LMS/HSS, with competitive results against Dilithium, especially in resource-limited IoT applications. SPHINCS+ is a stateless hash-based signature scheme standardized by NIST, prized for its conservative security assumptions but criticized for slow signing and large signatures. Picnic relies on zero-knowledge proofs and offers small public keys but large signatures. LMS/HSS is a lightweight hash-based scheme designed for constrained devices but requires careful state management. Dilithium, the NIST-selected lattice signature, offers strong performance but produces larger keys and signatures than FSDS-OTS+ under the reported benchmarks. The new scheme’s combination of small keys, small signatures, and fast verification positions it as an attractive option for next-generation blockchains, IoT networks, and quantum-era distributed systems.

One-time signatures occupy a special niche in post-quantum cryptography. Unlike multi-time schemes, they can sign only a single message per key pair, which simplifies security proofs and reduces the attack surface but requires key management strategies such as Merkle tree aggregation to scale. The FSDS-OTS+ design embraces this constraint and turns it into an advantage: the formal reduction proofs are cleaner because the adversary has fewer signing queries to exploit, and the homomorphic consistency checks allow verifiers to confirm that a signature was generated under the correct key without revealing the key itself. Zero-knowledge verifiability ensures that the verification process leaks nothing beyond the validity of the signature, a property that becomes increasingly important as signatures are deployed in privacy-sensitive contexts such as federated learning and threshold signing.

The work arrives at a pivotal moment for post-quantum cryptography standardization. NIST has been running an additional signature standardization process, with the second round of additional digital signature schemes documented in NIST IR 8610, reflecting the community’s recognition that the initially selected schemes may not cover all deployment scenarios. Lattice-based signatures for embedded IoT systems, post-quantum blockchain frameworks, and quantum-secured distributed ledgers have proliferated in recent literature, but few have combined formal QROM proofs with dataset-calibrated adversarial simulations. The FSDS-OTS+ authors argue that this dual assurance, pairing reductionist security guarantees with empirically grounded attack modeling, is what distinguishes a deployment-ready scheme from a theoretical curiosity.

The implications extend across the digital economy. Blockchains face an existential reckoning because their immutability means that transactions signed with quantum-vulnerable schemes cannot be retroactively protected; once a quantum computer arrives, every historical transaction on a vulnerable chain becomes a potential forgery target. IoT ecosystems, with billions of constrained devices, need signatures that fit within tight memory, bandwidth, and energy budgets while resisting quantum adversaries for decades of device lifetime. Decentralized finance platforms compound both problems, requiring high-throughput verification and long-term trust in immutable ledgers. FSDS-OTS+ addresses all three domains with a scheme that its authors position as secure-by-design, empirically confirmed, and deployment-ready. Whether it will be adopted by the broader community depends on independent scrutiny of its proofs and benchmarks, but the combination of formal QROM security, reduction robustness, and real-world performance numbers makes it a serious contender in the quantum-era cryptographic landscape.

Subject of Research: FSDS-OTS+: formal security proofs, reduction robustness, and quantum-adaptive improvements in the random oracle model

Article Title: FSDS-OTS+: formal security proofs, reduction robustness, and quantum-adaptive improvements in the random oracle model

Article References: Nauman, U., Hongyue, L., & Deng, M. (2026). FSDS-OTS+: formal security proofs, reduction robustness, and quantum-adaptive improvements in the random oracle model. Cluster Computing, 29(15), Article 862. https://doi.org/10.1007/s10586-026-06697-3

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06697-3

Keywords: FSDS-OTS, formal, security, proofs, reduction, robustness, quantum-adaptive, improvements, random, oracle, model, scientific research

The research, authored by Ruichao Xiao of Henan Industry and Trade Vocational College in Zhengzhou, China, builds on the observation that depth sensing has become a powerful ally for object detection. By pairing ordinary RGB images with depth maps, a system gains geometric structure information that helps it separate foreground objects from cluttered backgrounds. But existing RGB-depth salient object detection models, the paper argues, have largely focused on feature fusion while neglecting systematic adversarial defense—and even treat the RGB and depth channels independently when defending, ignoring cross-modal vulnerabilities that attackers can exploit.

At the heart of the new framework is a dual-stream encoder design. Two independent VGG-16 networks process the RGB image and the depth map, with the single-channel depth data normalized and replicated into three channels so it fits the architecture. From an input of 3 by 256 by 256 pixels, the encoders extract five layers of feature maps at different scales. The low-level layers carry spatial details such as textures and edges, while the higher layers encode global semantics like object categories and contours. The challenge, as in most multi-scale vision systems, is getting these levels to talk to each other effectively.

To solve that, the study introduces two fusion modules built around channel and spatial attention. The Enhanced Detail Integration Module handles the low-level features, applying a 3-by-3 convolution followed by channel attention and then spatial attention in series, which sharpens useful details while suppressing background noise. The Enhanced Semantic Integration Module works on the high-level features, and here the design is deliberately asymmetric: RGB features, rich in color and texture, are weighted first by channel attention, whereas depth features, whose value lies in spatial layout rather than channel distinction, are processed by spatial attention first. Reversing the attention order for the two streams matches each mechanism to the dominant property of its modality. A cross-layer transmission mechanism then feeds the fused low-level detail into the higher semantic layers, countering the dilution that plagues top-down feature propagation.

Downstream, an Enhanced Multi-scale Receptive Field Unit aggregates context across scales. It reduces channels with a 1-by-1 convolution, runs parallel convolutions with kernel sizes of 3, 5 and 7, applies dilated convolutions with hole rates of 1, 2 and 4 to widen the receptive field, and concatenates the results with a residual connection that preserves the original information. A densely connected pyramid decoder then fuses the enhanced features layer by layer before upsampling back to full resolution and producing the saliency map through a sigmoid activation.

The adversarial side of the framework is where the work becomes unusual. The author constructed adversarial samples using classic attack algorithms, including the Fast Gradient Sign Method, the multi-step Projected Gradient Descent, the Basic Iterative Method, the Carlini and Wagner attack, a multi-modal joint RGB-D attack, and a depth-specific attack designed to probe the depth channel’s vulnerability. Against this battery, the defense operates in two tiers. A passive front end applies WebP lossy compression to the RGB input, destroying small perturbations that are imperceptible to human vision while leaving the depth map uncompressed to protect spatial structure. An active component, the Contextual Correlations Aggregation Denoising Active Defense Module, is embedded at the low, middle and high layers of both encoders, where Gaussian denoising and attention-weighted context aggregation strip adversarial noise from the features themselves. Training combines a standard binary cross-entropy loss with a cross-modal defense enhancement loss weighted by a balance coefficient.

The experiments, run on an Ubuntu platform with an NVIDIA RTX 4090 GPU and trained with the Adam optimizer for 40 epochs at a batch size of 8, used three public benchmarks: NJU2K with 1,985 RGB-D image pairs, NLPR with 1,000 pairs at 640 by 480 resolution, and the 135-pair RGBD135 indoor dataset. On clean images the model reached a mean absolute error of 0.014 and an S-measure of 0.971 on NJU2K, beating the DMRA baseline’s 0.017 and 0.963, and posted an S-measure of 0.924 on NLPR, ahead of DPANet’s 0.915. It also recorded an MAE of 0.030 with an F-measure of 0.918 on NJU2K, surpassing recent state-of-the-art methods including D3Net, VST, SMTNet and SPNet. Precision-recall testing showed the advantage held across thresholds: at a recall of 0.8 on NJU2K, precision stayed at 0.945, compared with 0.912 for DPANet and 0.902 for DMRA.

Ablation studies confirmed that every component earns its place. Removing the two fusion modules produced the largest accuracy drop, with MAE rising from 0.032 to 0.051 on NJU2K, while removing cross-layer transmission or the multi-scale unit also degraded performance. Under PGD attack, the full robust model achieved the lowest MAE of 0.1635 and the highest F-measure of 0.5702, and dropping WebP compression alone cost 0.0349 in MAE and 0.0876 in F-measure—the single biggest hit—showing that the passive compression front end does real defensive work. Across FGSM, CW and BIM attacks, the ranking of variants stayed consistent: the combined WebP-plus-denoising configuration outperformed either mechanism alone by four to five percentage points, evidence that input-level and feature-level defenses are complementary rather than interchangeable.

Head-to-head against adversarially trained competitors, the framework held its lead. Under PGD it reached the best precision of 0.875, beating GACD at 0.862 and LeNo at 0.845; under the multi-modal joint attack it maintained 0.873; and under the depth-specific attack it hit 0.891 precision at a recall of 0.8, validating that the denoising module genuinely protects depth feature extraction even when the RGB channel is clean. On RGBD135 under attack, the model’s F-measure of 0.592 plus or minus 0.032 clearly outpaced GACD’s 0.519, and on NLPR its adversarial MAE of 0.098 plus or minus 0.006 beat LeNo’s 0.112. All results were averaged over three independent runs with 95 percent confidence intervals reported. Perhaps most striking for real-world use, the model runs at 62.5 frames per second—fast enough for video-rate applications—with moderate parameter and computational overhead.

The author is candid about limitations. The denoising module can blur high-frequency edge details, and the dense decoding mechanism makes the model computationally expensive. Future work, the paper suggests, could introduce an adaptive noise mechanism that dynamically adjusts defense intensity and explore lightweight network structures to shrink the parameter count. Even so, the study demonstrates that feature fusion and adversarial defense need not be separate research tracks: by organizing attention-ordered fusion, multi-scale context enhancement and a cascaded compression-denoising pipeline into one end-to-end framework, it offers a template for vision systems that must stay accurate not just in messy scenes, but in hostile ones—precisely the conditions autonomous vehicles and surveillance platforms face every day.

Subject of Research: RGB-depth salient object detection with multi-level feature fusion and adversarial defense

Article Title: RGB-depth salient object detection technology based on multi-level features and adversarial attacks

Article References: Xiao, R. (2026). RGB-depth salient object detection technology based on multi-level features and adversarial attacks. Discover Artificial Intelligence, 6(1), Article 1522. https://doi.org/10.1007/s44163-026-02508-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02508-1

Keywords: salient object detection, RGB-depth images, adversarial attacks, feature fusion, attention mechanisms, computer vision, WebP compression, deep learning, autonomous driving, robustness, NJU2K dataset, machine learning

The trustworthiness problem has two distinct faces, and the team led by Majid Ramedani and Martin Dyrba attacked both. The first is demographic confounding. Aging is the single greatest risk factor for Alzheimer’s disease, and the structural overlap between normal aging and pathology means a disease classifier may inadvertently rely on aging patterns rather than disease-specific atrophy. Brain age prediction models offer telling evidence of this: they typically estimate older ages for patients with neurodegenerative disease, a signature that the same confounding signal has leaked into the learned representations. The second face is unreliable probability estimates. The softmax layer in a convolutional neural network transforms raw output scores into a probability distribution over classes, but those disease likelihood scores indicate only the model’s relative confidence — not true statistical probabilities — and they come without confidence intervals. A model that says a patient has a 0.8 probability of Alzheimer’s is making a claim about its own internal scores, not about observed disease frequencies.

To test their solutions rigorously, the researchers built all their models on a shared 3D convolutional neural network architecture, ensuring that any performance differences could be attributed to the debiasing strategy rather than to architectural advantages. The baseline CNN consists of four convolutional blocks with 3×3×3 kernels, batch normalization, ReLU activation, and 2×2×2 max-pooling, followed by a bottleneck module that compresses whole-brain voxel-wise gray matter probability maps into a 4,096-dimensional feature vector. The inputs were whole-brain, voxel-wise gray matter maps derived from T1-weighted MRI scans, spatially normalized to MNI space and processed through a single standardized pipeline applied identically to all three cohorts. The models operated directly on whole-brain voxel-wise information rather than on predefined regional volumes, letting the networks find their own features in the raw gray matter signal.

The first debiasing approach, called CF-Net, is an adversarial confounder-free network built on a min–max optimization game. A feature extractor generates a feature vector that a classifier uses to predict diagnosis, while a separate confounder predictor attempts to predict the patient’s age from the very same features. The adversarial loss is defined using the squared Pearson correlation between true and predicted confounders, and minimizing it drives that correlation toward zero, encouraging statistical independence between the learned features and age. Training proceeds in three phases: the encoder and classifier train jointly first, then the confounder predictor activates while the others freeze, and finally a joint loop lets the encoder and confounder predictor interact to refine the representation. Learning rates follow an 8:1 ratio to balance the capacities of the 6.3-million-parameter classifier and the 0.8-million-parameter confounder predictor.

The second approach, PMDN — penalty-based metadata normalization — takes a fundamentally different route. Instead of playing an adversarial game, it treats the confounding effect as a general linear model, modeling the feature vector as f minus Mβ equals r, where M contains metadata, β are regression coefficients, and r is the residual, confounder-free feature vector. Because deep learning optimizes batch-wise, naive least-squares estimates of β would vary across batches and introduce instability. PMDN solves this by formulating a differentiable, bilevel optimization problem that jointly updates the network weights and the deconfounding parameters, with a penalty term that suppresses feature correlation with metadata. Alternating updates of the weights and β ensure stable convergence while the network learns task-relevant, residualized features. At inference, only the confounding components are subtracted from the 4,096-dimensional encoder features.

The results, tested on external cohorts after training on ADNI data, tell a striking story about stability. PMDN achieved the highest balanced accuracy at 0.71 with a standard deviation below 0.02 — roughly ten times lower than the baseline’s variability of about 0.11 and five times lower than CF-Net’s 0.06. All models posted comparable AUC values around 0.82, meaning their ability to discriminate diseased from healthy brains was essentially identical. But PMDN delivered the lowest Brier score at 0.22, against 0.32 for the baseline, indicating far more reliable probability estimates. The age bias quantification was even more revealing. On an age-balanced test set, the baseline model showed a consistent positive correlation between chronological age and predicted Alzheimer’s logits in cognitively normal subjects (r = 0.16), meaning older healthy people were systematically assigned higher disease risk. CF-Net barely moved that needle (r = 0.15) with greater variability, but PMDN reduced it to near zero (r = −0.07).

The mild cognitive impairment subgroup — the clinically critical category where age-related atrophy and pathological changes overlap most — showed the clearest separation between strategies. The baseline model exhibited a substantial age-logit correlation of r = 0.19, and CF-Net was actually worse at r = 0.20, suggesting adversarial training was ineffective at mitigating age bias in this population. PMDN achieved a near-zero correlation of r = 0.06, reflecting successful removal of age-confounded variance exactly where it matters most for early detection. In the diagnosed Alzheimer’s group, however, age-related associations persisted across all three models, with correlations around r = 0.15 unchanged by either debiasing method. The researchers attribute this to the greater overlap between age-related brain changes and disease-related neurodegeneration in established pathology, which makes it genuinely harder to disentangle the two signals. Even so, PMDN showed lower variability in the age-logit association for the Alzheimer’s group, indicating more consistent behavior across evaluation folds.

The second half of the study tackled the calibration problem with four post hoc methods applied to the CNN’s softmax outputs: two parametric approaches (logistic regression and probit regression) and two nonparametric ones (isotonic regression and Bayesian binning into quantiles, or BBQ). Logistic regression applies Bayes’ theorem under Gaussian class-conditional assumptions, reducing to an affine transformation of the raw scores, while probit regression substitutes the inverse normal link function. Isotonic regression fits a nondecreasing, piecewise-constant function using the pool-adjacent-violators algorithm, achieving an optimal linear-time solution without assuming any parametric form. BBQ divides outputs into bins based on empirical quantiles and estimates probabilities within each bin using a Bayesian beta-distribution approach, averaging over models with varying bin counts. The uncalibrated CNN outputs showed the largest calibration errors, with an expected calibration error of 0.28 and a maximum calibration error of 0.64 — meaning the model’s stated confidence could deviate from observed accuracy by nearly two-thirds at its worst.

Calibration transformed those numbers dramatically. Parametric methods produced moderate improvements, cutting the expected calibration error to roughly 0.13 and the maximum to about 0.29 with smooth, constrained curves. The nonparametric approaches did better still: isotonic regression and BBQ achieved expected calibration errors of 0.07–0.08 and maximum errors of 0.19–0.20, with isotonic also posting the lowest log-loss (0.50) and Brier score (0.16) while preserving the same F1-score of 0.73. Classification performance remained largely stable across all methods, confirming that calibration improves the reliability of predicted probabilities rather than the model’s discriminative ability. The researchers note that monotonic approaches produced smooth and reliable calibration curves, whereas the nonmonotonic BBQ method introduced irregularities — a reminder that flexibility in calibration comes at the cost of potential instability, and that evaluation should combine metrics like ECE and MCE with visual inspection of reliability plots.

The deeper lesson of the study is that deep learning models achieving similar classification performance can differ dramatically in calibration, stability, and fairness — dimensions that headline accuracy figures never capture. The contrast between the two debiasing strategies stems from their underlying optimization: PMDN applies a deterministic GLM-based correction that produces stable, consistent feature representations, while adversarial training relies on a min–max game with competing objectives that may converge incompletely, leaving residual confounding and explaining CF-Net’s variable performance. The findings carry real clinical weight for mild cognitive impairment populations, where models that lean on age as a proxy risk overestimating disease probability in older individuals. The team acknowledges limitations: age was the only confounder explicitly addressed, while sex, scanner variability, and site effects may also shape predictions; dataset size remains limited relative to the dimensionality of neuroimaging data; and adversarial training might reach its full potential with longer or more carefully tuned optimization. Future work will extend the approach to multiconfounder settings, larger and more diverse cohorts, and anatomically informed regional inputs such as hippocampal measures, which may encourage reliance on clinically established disease markers while reducing the influence of widespread age-related structural change.

Subject of Research: Artificial Intelligence

Article Title: Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration

Article References: Ramedani, M., Terli, J. C., Martensen, O. H., Strauß, T., Buerger, K., Dechent, P., Düzel, E., Fliessbach, K., Gemenetzi, M., Glanz, W., Hellmann-Regen, J., Hetzer, S., Incesoy, E. I., Janowitz, D., Kronmüller, M. T., Laske, C., Lüsebrink, F., Perneczky, R., Peters, O., … Dyrba, M. (2026). Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03805-x

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03805-x

Keywords: deep learning, Alzheimer's disease, machine learning, medical imaging, MRI, algorithmic fairness, calibration, neurodegenerative disease, Improving, fairness, robustness, deep

The shepherding problem, as the authors describe it, involves designing the movement law of steering agents, called shepherds, to navigate another set of agents, called sheep, which are driven by repulsive forces from the shepherds. The concept is bio-inspired, drawing on the way real sheepdogs herd livestock, and it has attracted attention for applications in robotics, group dynamics, and even nanochemistry, where active particles could potentially steer passive ones. Most existing algorithms, however, assume a central coordinator that can observe the entire system and communicate instructions to every shepherd, an assumption the Osaka team argues is not realistic enough for emerging applications where such infrastructure may be unavailable or unreliable.

To understand what makes the new approach different, it helps to look at the underlying sheep model. Each sheep agent moves in discrete time steps on a two-dimensional plane, its movement vector composed of five forces: separation, alignment, cohesion, repulsion from shepherds, and a random disturbance. The first three forces are a modified implementation of the classic Boid model introduced by Craig Reynolds in 1987, which has become a standard framework for simulating flocking behavior. Separation ensures agents do not collide, alignment steers them toward the average heading of local neighbors, and cohesion pulls them toward the average position of nearby agents. The repulsion force from shepherds uses a diverging potential function so that the closer a shepherd gets, the stronger the sheep’s urge to move away, which is precisely the mechanism that makes herding possible in the first place.

The shepherds themselves operate under strict information constraints. Each has a finite recognition range and can detect only its relative position to the goal, the relative positions of sheep and other shepherds within that range, and nothing else. The authors note that the first two types of sensing could be realized with proximity sensors, while goal-relative positioning could be achieved with recent methodologies such as single beacon-based positioning systems. Crucially, no shepherd ever needs to know what another shepherd intends or plans, which is what distinguishes this work from earlier decentralized algorithms that still implicitly required communication, such as one that needs shepherds to understand each other’s intentions or another that requires multiple rounds of messaging to reach consensus on which sub-flock each shepherd should chase.

The algorithm builds on an existing single-shepherd method called the Farthest-Agent Targeting algorithm, or FAT. In FAT, the shepherd chases the sheep farthest from the goal, maintaining an appropriate distance before pushing it toward the goal region. While effective for a single shepherd, FAT has two problems when generalized to multiple shepherds: all shepherds would target the same sheep, and the flock tends to scatter when a shepherd chases a sheep on the opposite side of the flock. The Osaka team’s extension introduces a parameter alpha that modifies target selection. Each shepherd chooses as its target the sheep maximizing a weighted difference between the sheep’s distance from the goal and its distance from the shepherd. When alpha is zero, the algorithm reduces to FAT. When alpha is large, each shepherd simply targets the closest sheep, preventing the scattering phenomenon. A moderate value balances both objectives.

Cooperation between shepherds emerges not from any messaging but from the spatial distribution induced by inter-shepherd repulsion built into the algorithm. Each shepherd’s movement vector is a weighted sum of four components: chasing its target sheep, maintaining an appropriate distance from neighboring sheep, pushing toward the goal region, and repelling other shepherds to avoid competition. The repulsion between shepherds is weighted by the shepherd’s distance from the goal, because shepherds need to be relatively close to each other at the final stage of navigation. The result is that although each shepherd independently chases its own target, the collective behavior resembles coordinated herding, a form of emergent cooperation arising purely from local rules and physical proximity.

The authors evaluated the algorithm through extensive numerical simulations across three initial flock configurations, a small flock of twenty sheep, a large flock of fifty, and two separate flocks of twenty and thirty, combined with three shepherd placement patterns. They compared the proposed method against FAT, FAT with occlusion, and an online-target switching algorithm. The proposed method achieved almost one hundred percent success rate regardless of the number of shepherds or placement pattern, while also outperforming the baselines in completion time and average path length. Both completion time and path length steadily decreased as more shepherds were added, suggesting stable and synergistic coordination. Notably, the online-target switching algorithm failed entirely at herding two separate flocks, while the proposed method handled them successfully.

The team also tested the algorithm under two additional parameter sets, one where sheep were more sensitive to shepherd movements and one where they were harder to navigate. The proposed method consistently showed higher success rates, lower completion times, and shorter path lengths than the baselines across all scenarios. A slight decrease in success rate appeared under the sensitive scenario, which the authors attributed to interference behaviors when multiple shepherds coincidentally chose the same target and drove it further from the goal, and to the difficulty of simultaneously containing all sheep within the goal region when the flock parameters made them more reactive.

Perhaps most striking was the comparison with centralized algorithms. The authors tested point-offset circling control, where shepherds form an arc formation directed by a central controller, and potential-based caging, where shepherds occupy the vertices of a regular polygon surrounding the flock. Both centralized methods showed increasing completion times and decreasing success rates as the number of shepherds grew, and neither could handle two separate flocks at all. The authors analyzed that the poor performance stemmed from the large interaction distance between shepherds and the flock, and that centralized algorithms were neither robust nor resilient enough to adapt to changes in flock shape caused by sheep movement. The communication-free approach, by contrast, thrived precisely where the centralized methods struggled.

The robustness and resilience tests round out the picture. When the shepherds’ sensing range was varied from fifty to four hundred units, success rates dropped only when the range was too short, suggesting the algorithm needs shepherds with reasonably long sensing capabilities but tolerates substantial variation otherwise. When shepherds were prone to random failures with probabilities up to forty percent per time step, with recovery probability of ten percent, the algorithm maintained high success rates, with completion times rising but traversal distances remaining almost unchanged. Even with additive sensing noise weighted at ten units, the algorithm tolerated relatively large sensing errors across all initial configurations. The authors attribute this fault tolerance to the decentralized mechanism itself, since no single shepherd is indispensable and each operates independently based on its own observations. Future work, they note, includes investigating whether the communication-free coordination mechanism can extend to other navigation tasks and establishing mathematical properties such as stability and convergence, which remain challenging due to the nonlinearity of the underlying swarm model.

Subject of Research: Communication-free decentralized shepherding navigation of passive agent flocks with multiple steering agents

Article Title: Communication-free shepherding navigation with multiple steering agents

Article References: Li, A., Ogura, M., & Wakamiya, N. (2023). Communication-free shepherding navigation with multiple steering agents. Frontiers in Control Engineering, 4, Article 989232. https://doi.org/10.3389/fcteg.2023.989232

Image Credits: AI Generated

DOI: 10.3389/fcteg.2023.989232

Keywords: shepherding, swarm guidance, decentralized control, multi-agent systems, flocking, bio-inspired robotics, communication-free navigation, emergent coordination, robustness, resilience, control engineering, agent-based modeling

The distinction between symmetric and asymmetric noise, which the experimental design explicitly separates, reflects two very different real-world failure modes. Symmetric noise flips a label to any other class with equal probability, which approximates random annotation errors such as accidental clicks in a labeling interface. Asymmetric noise, by contrast, systematically confuses certain class pairs, mirroring the way human annotators genuinely mistake visually similar categories, for example confusing one endoscopic finding with another that shares its appearance. Asymmetric noise is widely regarded as the harder case because the errors are correlated with the input features themselves, so a model cannot easily separate signal from corruption. The observation that the optimal reward weight depends on both the noise type and the noise rate suggests that practitioners deploying this method should treat the scaling parameter as a tunable diagnostic rather than a fixed constant, adjusting it based on what is known or suspected about the annotation process that produced the training data.

The reward terms at the heart of the method draw on properties that deep networks are known to exhibit early in training. Empirical studies of the learning dynamics of neural networks trained on corrupted labels have repeatedly shown that networks first fit the patterns that generalize before beginning to memorize label errors, a phenomenon sometimes described as the early-learning phase. During this window, predictions on correctly labeled samples tend to be confident and consistent, while predictions on mislabeled samples are often less coherent. Confidence, similarity, and diversity measured across training batches can exploit this asymmetry: samples whose predictions align with the emerging consensus of the batch receive relatively more weight, while samples that conflict with it are down-weighted, which is precisely the behavior needed to reduce the influence of erroneous labels without discarding the informative ones.

Computing these rewards on batches rather than on individual examples is a design choice with practical consequences. Batch-level statistics act as a form of implicit ensemble, since the aggregated behavior of many samples provides a more stable reference than any single prediction. This also means the method integrates naturally into standard mini-batch training pipelines without requiring a second model, a separate noise-estimation stage, or a preliminary filtering pass over the dataset. Approaches that rely on two networks trained in parallel, on small-loss sample selection, or on iterative label correction each add engineering complexity and computational cost; a modified loss function preserves the simplicity of the standard training loop, which matters for teams scaling training to larger datasets or tighter compute budgets.

The baselines chosen for comparison span a meaningful slice of the noisy-label loss literature. Mean absolute error is theoretically noise-tolerant under certain conditions but is notorious for slow convergence and poor performance on hard datasets, because its gradients do not grow with the size of the error. Generalized cross entropy interpolates between cross entropy and mean absolute error through a single parameter, trading off the fast convergence of the former against the noise tolerance of the latter. Normalized cross entropy rescales the loss to counteract the tendency of cross entropy to keep pushing already-confident predictions. By outperforming both the plain cross entropy baseline and these established robust alternatives, particularly at moderate-to-high noise rates, the reported results place RECE within a competitive landscape rather than against strawmen, though the comparison remains confined to parameter-light loss modifications rather than the heavier algorithmic machinery of co-teaching, mentor networks, or label-correction pipelines.

The choice of benchmarks also deserves attention. MNIST and Fashion-MNIST are small, low-resolution datasets where even simple architectures achieve high accuracy on clean data, so they primarily serve to illustrate behavior under controlled corruption rather than to predict performance on production workloads. CIFAR-10, with its greater intra-class variability, is a more demanding and more representative test. The inclusion of an endoscopy dataset is arguably the most consequential, because it grounds the evaluation in a domain where label noise is not a laboratory artifact but a persistent operational reality, and where the cost of a model that has memorized erroneous annotations shows up in clinical decision support. Robustness results that hold across this heterogeneous set, spanning toy, standard, and domain-specific data, carry more weight than results on any single benchmark.

The unified scaling parameter also connects the method to a recurring theme in the robust-learning literature: the trade-off between fitting the training signal and resisting corruption. When the reward weight is small, the objective behaves close to ordinary cross entropy, retaining its strong discriminative drive but little protection against noise. When it is large, the reward terms dominate, and the model risks underfitting the genuinely informative part of the data. The finding that the best setting shifts with the noise regime is consistent with theoretical analyses showing that no single loss is uniformly noise-tolerant across all noise models; robustness must be purchased with some loss of fitting power, and the exchange rate depends on how much corruption is present. This framing suggests a practical workflow in which a small validation set with verified clean labels is used to tune the parameter, which is a far cheaper requirement than cleaning the entire training set.

It is worth noting what the loss-function approach does and does not address. Label noise in deployed systems is not always homogeneous: some errors are random flips, some are systematic confusions, and some arise from ambiguous instances that even experts would label differently. Loss modification operates at the level of gradient weighting, gently steering optimization away from samples whose labels conflict with the model’s learned structure, but it cannot recover the true label of a corrupted example, nor can it resolve ambiguity in the underlying data. Complementary techniques such as noise-transition matrices, sample filtering, and label correction attack the problem from other angles, and hybrid systems that combine a robust loss with a lightweight correction stage are a natural direction for future work. The evidence here establishes that reward-augmented cross entropy is a strong and stable foundation for such systems rather than a complete solution on its own.

Finally, the emphasis on stability across noise levels addresses a failure mode that practitioners know well: methods that excel at one noise rate sometimes degrade sharply at another, or interact badly with standard hyperparameters such as learning rate schedules and augmentation strength. A loss that degrades gracefully, maintaining solid performance from clean data through moderate and high corruption, reduces the risk of silent failure when the true noise characteristics of a dataset are unknown. Combined with the modest implementation overhead of adding reward terms and one scaling parameter to an existing cross entropy objective, this positions the approach as a pragmatic default for teams training classifiers on imperfectly labeled data, particularly in applied fields where annotation quality cannot be guaranteed in advance.

Subject of Research: RECE Loss: A Reward-Enhanced Cross Entropy Loss Function for Learning with Noisy Labels

Article Title: RECE Loss: A Reward-Enhanced Cross Entropy Loss Function for Learning with Noisy Labels

Article References: Nguyen, T. T. H., Nguyen, D. T., Nguyen, T. T., & Vu, H. (2026). RECE Loss: A Reward-Enhanced Cross Entropy Loss Function for Learning with Noisy Labels. Neural Computing and Applications, 38(20), Article 810. https://doi.org/10.1007/s00521-026-12500-6

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12500-6

Keywords: machine learning, noisy labels, loss functions, deep learning, cross entropy, robustness, neural networks, label noise, RECE loss, classification, RECE, Loss

The study, led by Syed Shuja Hussain of Universiti Malaysia Pahang Al-Sultan Abdullah with colleagues at Universiti Teknologi PETRONAS and other institutions, took a deliberately skeptical approach. Rather than asking which architecture reports the highest accuracy, the team conducted a PRISMA-guided systematic review of deep learning-based Android malware detection studies published between 2021 and 2026, then complemented the review with a controlled reliability audit of representative CNN, GRU, Transformer, and Hybrid CNN-LSTM models on the CICMalDroid 2020 dataset.

The review began with 739 records retrieved from IEEE Xplore, ACM Digital Library, Scopus, Web of Science, ScienceDirect, Springer, Wiley, and Google Scholar. After removing 214 duplicates and screening out 421 records that fell outside the Android malware detection scope or lacked deep learning components, 104 full-text articles were assessed for eligibility. Forty-three were excluded for insufficient Android focus, missing deep learning detection components, or weak empirical detail. The final synthesis retained 61 studies, including 33 primary technical studies that formed the architecture-level evidence base.

That evidence base is architecturally diverse. CNN and image-based models dominated the landscape, appearing in 51.5 percent of primary technical studies, typically learning from DEX images, bytecode images, opcode images, manifest features, or permission encodings. Sequential models such as LSTM, GRU, and BiGRU accounted for 24.2 percent, processing API-call sequences, system-call traces, and network flows. Transformer and attention-based models made up 21.2 percent, tokenizing traffic, opcode, API, or DEX representations to capture long-range contextual dependencies. Graph neural networks, autoencoders, hybrid multimodal systems, lightweight deployment-oriented models, and explainable AI-oriented architectures rounded out the field.

But the review’s central finding is one of caution rather than celebration. When the team scored 28 non-redundant representative studies across nine evidence dimensions, including dataset transparency, artifact clarity, architecture clarity, metric adequacy, validation realism, robustness evidence, explainability grounding, reproducibility support, and deployment evidence, the aggregate pattern revealed that architectural innovation is running ahead of evidence maturity. Metric adequacy was consistently reported across all 28 studies, and dataset transparency, artifact clarity, and architecture clarity were generally well represented. Yet only 2 of 28 studies provided clear robustness or evasion evidence, only 5 offered clear explainability evidence mapped to Android artifacts, and none provided complete reproducibility support. Deployment evidence was clearly reported in just 5 studies.

The controlled empirical validation on CICMalDroid 2020 reinforced this argument. All four representative architectures achieved competitive clean cross-validation performance, with macro F1-scores between 0.895 and 0.955. The Hybrid CNN-LSTM achieved the best clean profile with an accuracy of 0.945 ± 0.004 and macro F1-score of 0.955, followed by the Transformer at 0.928 ± 0.006 accuracy and 0.925 macro F1, the GRU at 0.907 ± 0.008 and 0.910, and the CNN at 0.911 ± 0.005 and 0.895. But clean accuracy was only the starting point of the audit.

Class-wise analysis revealed that adware remained the most challenging category for every architecture evaluated, even when aggregate performance looked strong. The malware false-negative rate, a security-critical metric measuring undetected malicious applications, ranged from 0.054 for the Hybrid CNN-LSTM to 0.119 for the CNN. Repeated-run stability across multiple random seeds showed the Hybrid CNN-LSTM achieving the highest mean macro F1-score of 0.955 with the least variation, while the CNN, GRU, and Transformer showed progressively lower stability. Multi-level test-time feature masking at 5, 10, 20, and 30 percent perturbation levels exposed architecture-dependent robustness differences invisible to aggregate accuracy. Under 30 percent masking, the CNN’s macro F1 dropped from 0.895 to 0.706, the GRU fell from 0.910 to 0.744, the Transformer declined from 0.925 to 0.768, and the Hybrid CNN-LSTM retained the best masked performance at 0.814 from 0.955.

The study proposes an architecture-artifact-validation triad as the correct unit of evaluation, arguing that a CNN trained on DEX images, a GRU trained on API-call sequences, a Transformer trained on traffic tokens, and a graph neural network trained on function-call graphs are not simply different models for the same evidence but fundamentally different evidence systems with distinct failure modes. Image-based models may learn visual separability without preserving meaningful malware semantics. Sequential models depend on trace completeness and whether malicious behavior is actually triggered during observation. Graph-based evidence relies on fragile call-graph extraction vulnerable to reflection, dynamic loading, and obfuscation. Attention maps and saliency outputs are useful only when mapped back to verifiable Android artifacts such as APIs, permissions, DEX methods, or manifest elements.

The authors introduce a claim-aligned reliability audit that matches each layer of evaluation to the type of claim it can legitimately support. Clean accuracy authorizes only baseline benchmark discrimination. Class-wise F1-scores and malware recall authorize threat-category reliability. Robustness claims require perturbation, obfuscation, adversarial, temporal, or cross-dataset evidence reported separately. Deployment claims require model size, inference latency, memory footprint, preprocessing cost, hardware setting, energy impact, and real-device or edge feasibility. The audit also subjects the study’s own empirical validation to the same standard, acknowledging that it uses a single public dataset, a shared hyperparameter protocol, and feature-level perturbation rather than APK-level obfuscation or real-world adversarial transformation.

The findings support a transition from accuracy-centric benchmarking to reliability-centric evaluation, where deep learning-based Android malware detection is assessed using credible datasets, realistic validation protocols, robustness testing, artifact-grounded interpretability, reproducibility assets, and operational feasibility evidence. The authors argue that future studies should report class-wise metrics, false-negative analysis, repeated-run stability, temporal or cross-dataset validation, adversarial or obfuscation-aware testing, transparent implementation details, and deployment-relevant evaluation rather than collapsing results into a single headline accuracy score.

Subject of Research: Deep learning Android malware detection under realistic evaluation constraints

Article Title: Deep learning-based android malware detection under realistic evaluation constraints: an architecture-centric review and reliability-aware empirical validation

Article References: Hussain, S. S., Razak, M. F. A., Ismail, I., Firdaus, A., Azeem, A., Zara, J., & Akhter, S. (2026). Deep learning-based android malware detection under realistic evaluation constraints: an architecture-centric review and reliability-aware empirical validation. Frontiers in Artificial Intelligence, 9, Article 1964400. https://doi.org/10.3389/frai.2026.1964400

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1964400

Keywords: Android malware detection, deep learning, reliability audit, systematic review, CICMalDroid 2020, robustness, explainable AI, Deep, learning-based, android, malware, detection

The research, conducted by Shuyan Li, Jinge Yao, Xina Ji, Mingjie Fang, and Qiwei Pang, surveyed 214 Chinese firms deeply involved in the international seaborne import of bulk commodities, including energy resources, mineral products, and agricultural goods. Using a two-wave survey design separated by roughly three months and structural equation modeling, the team traced how three managerial processes—resource acquisition, resource bundling, and resource leveraging—feed into data analytics capability, which in turn supports two distinct disruption-response capabilities: supply chain robustness and supply chain resilience.

The theoretical foundation combines two complementary frameworks. Resource orchestration theory holds that possessing valuable resources is not enough; firms must actively structure, bundle, and leverage them to build capabilities. Organizational information processing theory adds that as environmental uncertainty rises, so do the demands on a firm to collect, interpret, and apply complex information. In bulk commodity maritime supply chains, where critical resources are dispersed across overseas suppliers, shipping companies, ports, logistics providers, and internal functions—many controlled by independent organizations—both theories point to the same conclusion: the ability to manage fragmented resources and fragmented information together determines whether a firm can withstand disruption.

The empirical results confirm a semi-sequential logic. Resource acquisition was positively associated with resource bundling (standardized β = 0.345, p < 0.001), and resource bundling was positively associated with resource leveraging (β = 0.354, p < 0.001). In other words, firms that secure access to transportation capacity, port facilities, sourcing channels, digital technologies, and professional expertise are better positioned to integrate those resources into coherent capabilities, and firms with well-integrated capabilities are better able to mobilize and deploy them in actual operations such as supplier assessment, route planning, port selection, and inventory adjustment.

All three orchestration processes were also positively associated with data analytics capability, but with striking differences in strength. Resource acquisition showed the strongest association (β = 0.422, p < 0.001), followed by resource leveraging (β = 0.233, p < 0.001) and resource bundling (β = 0.177, p = 0.010). Together, the three processes explained 36.8 percent of the variance in analytics capability. This pattern suggests that access to rich, timely information—vessel movements, freight rates, port congestion, weather conditions, overseas supply conditions—is the single most important ingredient in building the capacity to turn heterogeneous data into predictive and prescriptive insights.

Data analytics capability, in turn, proved to be a powerful predictor of both disruption-response outcomes. Its association with maritime supply chain robustness—the ability to maintain acceptable functioning during disruption—was strong (β = 0.529, p < 0.001), and its association with resilience—the ability to adapt, respond, and recover after disruption—was similarly strong (β = 0.514, p < 0.001). The model explained 28.6 percent of the variance in robustness and 27.0 percent in resilience. Notably, control variables such as firm size, total sales, and firm age showed no significant associations with either outcome, indicating that these relationships are not simply artifacts of larger or older firms being better prepared.

Bootstrap analysis with 5,000 resamples confirmed significant indirect effects: resource acquisition, bundling, and leveraging each influenced robustness and resilience through data analytics capability rather than through direct paths. This finding is theoretically important because it positions analytics as the information-processing mechanism that converts orchestrated resources into disruption-response capability. It also reinforces the empirical distinctiveness of robustness and resilience, which the study measured separately and found to be correlated but not interchangeable (HTMT = 0.469, well below the conservative threshold of 0.85).

The practical implications are substantial. Managers should treat resource acquisition, bundling, and leveraging as an integrated process rather than as separate activities, because obtaining resources alone does not strengthen disruption management unless those resources are integrated and deployed. Investment in analytics should extend beyond software to include the information sources, human expertise, and external relationships that make analytical tools useful. Firms should also embed analytical outputs into routine decisions about supplier selection, transportation capacity, route choice, discharge ports, shipping schedules, and contingency planning, rather than leaving them isolated within technical functions.

The study’s authors acknowledge limitations, including the focus on Chinese firms, the use of a single key informant per firm, and the concurrent measurement of analytics capability, robustness, and resilience at Time 2, which limits strong causal interpretation. Future research with multi-informant designs, objective performance measures, and longitudinal data could strengthen causal identification and test whether these mechanisms vary across different maritime economies and commodity sectors.

Nevertheless, the findings offer a clear message for firms dependent on seaborne bulk commodity imports: building robust and resilient supply chains requires more than accumulating resources or purchasing digital technologies. It requires connecting external resource acquisition with integration and operational deployment, while embedding data analytics into sourcing, shipping, port, inventory, and disruption-response decisions. In an era of climate extremes, geopolitical tension, and cascading trade network risks, that integrated capability may be the difference between a supply chain that buckles and one that bends, adapts, and recovers.

The distinction between robustness and resilience that underpins the study has deep roots in systems engineering and ecology before it migrated into supply chain management. In engineering contexts, robustness describes a system’s capacity to preserve performance when parameters change, while resilience in ecological science describes how a system absorbs disturbance and reorganizes while retaining function. Supply chain scholars adapted both ideas to capture a temporal asymmetry: robustness operates during a disruption, resilience after it. For a bulk commodity importer, this means robustness might involve maintaining delivery schedules when a typhoon delays vessels, whereas resilience might involve rebuilding procurement flows weeks later after a geopolitical rupture has severed a sourcing channel entirely. The finding that analytics capability predicts both outcomes with similar strength suggests that a single information-processing foundation can support what are otherwise organizationally distinct activities—static protection and dynamic reconfiguration.

Bulk commodities provide a particularly demanding setting for testing these ideas. Unlike containerized manufactures, bulk cargoes such as iron ore, coal, grain, and crude oil move in large, relatively homogeneous volumes through specialized ports and vessel classes, which means substitution across routes or carriers is often constrained by infrastructure rather than by demand. The geographic concentration of production and dispersion of consumption that characterizes many bulk commodities creates long, thin supply chains in which a single chokepoint—a mine outage, a canal blockage, a terminal closure—can propagate widely. This structural fragility makes the sector a useful natural laboratory for studying disruption response, and it explains why the authors chose firms substantially involved in seaborne imports rather than a general sample of traders.

The two-wave survey design deserves attention as a methodological safeguard. Much supply chain research collects all measures from a single questionnaire at a single moment, which invites common method bias—the risk that respondents’ general optimism inflates apparent relationships among variables. By measuring the resource orchestration processes several months before measuring analytics capability, robustness, and resilience, the study at least partially separates the predictor and outcome measurements in time, reducing this concern. Structural equation modeling then allowed the authors to estimate an entire network of relationships simultaneously, testing whether the data were consistent with a sequential process running from acquisition through bundling to leveraging, rather than evaluating each link in isolation.

Bootstrap analysis, used to test the indirect effects, is a resampling technique that repeatedly draws new samples from the observed data to estimate the uncertainty of a statistic without assuming a normal distribution. In mediation testing, this matters because indirect effects—the product of two pathway coefficients—often have skewed distributions that classical significance tests handle poorly. Five thousand resamples provide a reasonably stable empirical distribution from which confidence intervals can be constructed, giving the mediation findings more credibility than a simple significance test of individual paths would.

The theoretical integration is also noteworthy. Resource orchestration theory emerged from strategic management research on how managerial actions convert resource ownership into competitive advantage, while organizational information processing theory traces to classic work arguing that organizational design should match the information demands of the task environment. Combining them yields a specific, testable proposition: resources dispersed across organizational boundaries generate potential only when firms can process the information those resources produce. The empirical pattern fits this proposition. The strongest pathway ran from resource acquisition to analytics capability, consistent with the idea that analytics is fundamentally hungry for input data—vessel positions, freight indices, weather feeds, supplier signals—and that firms lacking access to such feeds have little to analyze regardless of their software investments.

Why resource bundling showed the weakest association with analytics capability invites further inquiry. One plausible interpretation is that bundling is an integrative activity whose value is realized downstream, in leveraging and in operational execution, rather than in information generation. Another is that bundling’s information content is largely internal—coordination routines, shared platforms, consolidated inventories—whereas the analytics capability measured here appears oriented toward external environmental scanning. Disentangling internal from external information processing could be a productive direction for subsequent work.

The study’s setting in China carries both advantages and limits for generalization. China is the world’s largest importer of many bulk commodities, so Chinese importers face acute exposure to maritime chokepoints and export-country concentration, making the context substantively important rather than merely convenient. At the same time, Chinese firms operate within a distinctive logistics infrastructure environment, including extensive state investment in port capacity and digital trade platforms, and within unique regulatory and financing conditions. Whether the same relationships hold for importers in other maritime economies, for exporters rather than importers, or for containerized rather than bulk trade remains an open empirical question.

The absence of significant effects for firm size, sales, and age is itself informative. If larger or longer-established firms were simply better at disruption management, the analytics pathway might be a proxy for organizational scale. The pattern instead suggests that capability-building processes are available to firms of varied profiles, which carries practical weight for mid-sized importers who cannot match the resource endowments of global majors but can still invest in acquisition relationships and analytical competence.

For policy observers, the findings also speak to broader debates about supply chain security. Governments increasingly emphasize reshoring, stockpiling, and diversification as responses to commodity supply risk. The study’s evidence suggests a complementary firm-level lever: because resilience depends on rapid reconfiguration of suppliers, routes, ports, and carriers, the private capability to sense and respond may amplify the value of any public diversification measure. A diversified sourcing base without the informational capacity to recognize when to switch is arguably as vulnerable as a concentrated one.

Finally, the study contributes to a maturing conversation about data analytics in supply chains by specifying a mechanism rather than merely documenting an association. Much prior work has asked whether big data analytics improves resilience; this research asks where analytics capability comes from and shows it arises through identifiable managerial processes. That reframing shifts the practical question from whether to adopt analytics to how to build the resource base and orchestration routines that make analytics consequential when disruption arrives.

Subject of Research: The role of data analytics capability in linking resource orchestration to robustness and resilience in bulk commodity maritime supply chains

Article Title: From resource orchestration to robustness and resilience: the role of data analytics capability in bulk commodity maritime supply chains

Article References: Li, S., Yao, J., Ji, X., Fang, M., & Pang, Q. (2026). From resource orchestration to robustness and resilience: the role of data analytics capability in bulk commodity maritime supply chains. Frontiers in Marine Science, 13, Article 1969423. https://doi.org/10.3389/fmars.2026.1969423

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1969423

Keywords: maritime supply chains, data analytics capability, resource orchestration, supply chain resilience, supply chain robustness, bulk commodities, organizational information processing theory, resource, orchestration, robustness, resilience, role