The research, led by Xinrong Sun of Shandong University together with Yunting Tao of Fudan University and Binzhou Polytechnic, Fanyu Kong and Guoyan Zhang of Shandong University, tackles a dilemma that has grown sharper as machine learning has colonized remote sensing. The state-of-the-art method for segmenting hyper-spectral images, known as joint sparse coding-based clustering, or JSCC, produces remarkably accurate maps of ground targets. But its core phases, dictionary construction and joint sparse recovery, are dominated by large-scale matrix multiplications and matrix pseudo-inversions, operations whose cost explodes as image size and spectral band count grow. For a resource-constrained client, outsourcing those computations to a powerful cloud server is the obvious move, and it is also a dangerous one.

The danger is straightforward: hyper-spectral imagery of a military installation, a disaster zone, or a commercially sensitive mining site is not public data. Handing raw matrices to an untrusted cloud server exposes them to curious operators, lazy servers that might return fabricated results to save money, and outright malicious adversaries who try to reconstruct the original imagery from whatever they observe. Existing cryptographic defenses each carry their own burdens. Secure multi-party computation requires elaborate interactive protocols and numerous secure multiplications. Fully homomorphic encryption inflates data into ciphertexts many times larger than the plaintext and relies on depth-consuming iterative approximations even for something as basic as matrix inversion. For high-dimensional matrix workloads, both approaches can cost more than they save.

That is why the field has long favored a lighter technique called matrix blinding, in which the client disguises its data with secret transformation matrices before sending them out. The trouble with previous blinding schemes, the authors argue, is the secret key itself. To encrypt a data matrix, earlier methods needed at least two large sparse key matrices, one for each dimension, and storing them consumed significant client-side resources. Those storage demands limited how widely the techniques could be deployed and, ultimately, how well the underlying analysis performed. The new work replaces those bulky key matrices with something far leaner: compact index sets that behave like matrices without ever being stored as matrices.

The heart of the scheme is a novel matrix encryption method built from three index sets. Each set contains a random permutation index, its inverse, and a value index of coefficients drawn from randomly generated two-by-two orthogonal transformations. Together these indices let the client perform elementary transformations on the rows and columns of a sensitive matrix, permuting them, scaling them, and mixing adjacent rows or columns, entirely through element-wise arithmetic. Because the orthogonal coefficients satisfy a normalization condition, the transformations are perfectly invertible: applying the corresponding inverse encryption restores the original matrix exactly. Crucially, the index sets are single-use, generated fresh for each encryption task, so no adversary can ever observe two different matrices scrambled by the same key.

The elegance of the design lies in a set of algebraic properties the authors prove formally. Encrypting the columns of one matrix and multiplying it by another is equivalent to multiplying the original matrix by a row-encrypted version of its partner. Encrypting a product is equivalent to encrypting one of its factors. Transposing an encrypted matrix equals encrypting the transposed matrix, and inverting an encrypted matrix equals encrypting the inverse. These associativity, transposition, and inversion properties mean the cloud server can perform ordinary matrix multiplication and pseudo-inversion on the blinded inputs, and the client can decrypt the blinded output to recover exactly the result it would have obtained by computing in the clear. The server learns nothing, because the blinded matrices are computationally indistinguishable from random matrices filled with uniformly distributed noise, a property the authors establish through a formal indistinguishability proof.

Security against cheating is handled by a sampling-based verification method. A lazy or malicious server might return a plausible-looking but incorrect matrix, so the client checks randomly selected columns of the returned result against the encrypted inputs using lightweight element-wise computations, with fresh random weights generated for every verification round. The authors show that any incorrect result slips past a single round of checks with probability at most one half, so after twenty rounds, the setting used in their experiments, the chance of a forged result being accepted falls below one in a million. Verification, like encryption and decryption, never requires the client to touch a full matrix operation.

The performance numbers are striking. Across simulated matrix datasets ranging from modest to very large scales, the new scheme outperformed the leading matrix-blinding competitors by 4.15 to 10.79 percent on average, and beat homomorphic-encryption and multi-party-computation baselines by wider margins on multiplication tasks. Theoretically, the scheme cuts the cost of a matrix multiplication from cubic complexity to a quadratic form, and slashes matrix pseudo-inversion from cubic to linear in the matrix dimensions. When the full JSCC segmentation pipeline was run on five real hyper-spectral datasets, including the well-known Indian Pines, Salinas, Botswana, and Pavia scenes, the outsourced version completed the analysis 73.49 to 80.45 percent faster than the original algorithm, while producing segmentation maps indistinguishable in quality from those computed locally. Numerical errors introduced by the encryption and decryption round-trips were below ten to the minus fourteenth, negligible against the precision of the underlying data.

The team also stress-tested the blinding method itself by building a neural network inversion attacker, a two-stream convolutional and up-convolutional model trained on pairs of original and blinded super-pixel matrices, inspired by techniques for inverting visual representations. Because every matrix is scrambled with independently generated index sets, the attacker could never learn a general inverse mapping. On held-out images the reconstructed outputs showed mean squared errors of roughly 0.026 to 0.052, peak signal-to-noise ratios of only about 13 to 16 decibels, and spectral angle deviations of 23 to 34 degrees, meaning the recovered data lost both fine spatial texture and the spectral direction of the original pixels. In plain terms, the neural network produced blurry, spectrally distorted ghosts rather than usable imagery.

The implications reach beyond satellite imagery. Matrix multiplication and pseudo-inversion sit at the core of many machine learning algorithms, and the authors note that their blinding method applies to any workload dominated by those operations, from K-means clustering to dimensionality reduction, and could support the linear layers of deep networks when combined with secure protocols for non-linear operations. As hyper-spectral sensors proliferate on drones, small satellites, and ground platforms, and as privacy regulation tightens around geospatial data, the ability to rent cloud-scale computation without surrendering cloud-scale secrets may determine who gets to turn raw spectral light into actionable knowledge. This study suggests the key to that future may be nothing more than a handful of cleverly shuffled indices.

Subject of Research: Privacy-preserving cloud outsourcing of hyper-spectral remote sensing image analysis using matrix blinding with index-set keys

Article Title: A privacy-preserving hyper-spectral remote sensing image analysis framework based on matrix outsourcing computation

Article References: Sun, X., Tao, Y., Kong, F., & Zhang, G. (2026). A privacy-preserving hyper-spectral remote sensing image analysis framework based on matrix outsourcing computation. Cybersecurity, 9(1), Article 225. https://doi.org/10.1186/s42400-026-00667-3

Image Credits: AI Generated

DOI: 10.1186/s42400-026-00667-3

Keywords: hyper-spectral imaging, remote sensing, cloud computing, privacy-preserving computation, matrix blinding, matrix outsourcing, secure outsourcing, joint sparse coding, image segmentation, data privacy, cryptography, machine learning

A team at Jiangnan University’s School of Intelligent Manufacturing in Wuxi, China, has now proposed a way to soften that conflict at an unusually fine level of granularity. In a study published in Cluster Computing, Ziyi Zhao, Lei Su, and Ke Li introduce pFedDS, short for parameter domain sensitivity based personalized federated learning. Rather than treating a neural network as an indivisible bundle of weights to be averaged, their method interrogates each individual parameter, asking how strongly it reacts to the idiosyncrasies of a given client’s data, and then adjusts how that parameter is trained accordingly. The result, they report, is a measurable accuracy gain over eight state-of-the-art competing methods on two real industrial datasets.

To understand why this matters, it helps to consider how earlier attempts at personalization have worked. Many personalized federated learning schemes slice the model into two fixed territories: a shared portion, typically the lower layers that capture general features, and a personalized portion, often the upper layers or a dedicated head, which each client trains freely on its own data. The split is decided in advance and applied uniformly to every participant. The Jiangnan University authors argue that this static partitioning neglects the actual effects of data heterogeneity. A layer that is safely generic for one factory’s camera images may be deeply entangled with local quirks for another, and a one-size-fits-all boundary between shared and private knowledge cannot capture that variation.

The core innovation of pFedDS lies in how it measures sensitivity. During training, the method tracks each parameter’s deviation from the global update direction, the trajectory that the federated model as a whole would naturally follow. Parameters whose local updates stray far from that collective path are judged highly sensitive to the non-IID character of the client’s data. Those that hew closely to the global direction are deemed insensitive, meaning they encode knowledge that generalizes across participants. Crucially, this measurement is dynamic: it is recomputed as training proceeds, so the sensitivity profile of the model evolves alongside the data distribution and the optimization landscape rather than being frozen at initialization.

Once sensitivity has been quantified, pFedDS acts on it through learning rates. During local training, each parameter is updated with a sensitivity-aware step size. High-sensitivity parameters, the ones most entangled with local data peculiarities, receive larger learning rates, allowing them to drift toward local knowledge and specialize for the client at hand. Low-sensitivity parameters receive smaller learning rates, which restrains their movement and preserves the globally useful knowledge they carry. In effect, the method performs a continuous, per-parameter negotiation between global and local interests, replacing the blunt binary of shared versus personalized layers with a fine-grained spectrum in which every weight finds its own equilibrium point.

This design philosophy connects to a broader trend in the federated learning literature. The field has explored neuron-wise learning rates, parameter importance estimation borrowed from network pruning techniques, distribution-aware sub-model extraction, and adaptive local aggregation schemes, all in an effort to reconcile collective training with individual variation. What distinguishes pFedDS is the specific signal it uses, deviation from the global update direction, and the way that signal directly modulates optimization dynamics rather than merely deciding which parts of the model to share. The approach aims at adaptive knowledge fusion, letting global and local information blend in proportions that differ from parameter to parameter and from client to client.

The empirical case for the method rests on two practical industrial datasets rather than synthetic benchmarks. The first is a precision component defect dataset collected by the authors themselves, and the second is the Northeast University defect dataset, a surface defect collection widely used in industrial inspection research. Defect detection is a demanding testbed for federated learning because production lines differ enormously: lighting conditions, camera angles, material finishes, and the very types of flaws that appear vary from plant to plant, producing exactly the kind of severe data heterogeneity that breaks naive federated averaging.

Across these datasets, pFedDS delivered consistent improvements. Under pathological distribution settings, an extreme form of non-IID data in which each client holds samples from only a limited set of classes, the method achieved test accuracy between 1.34 and 1.4 percentage points higher than eight state-of-the-art baselines. Under extremely Dirichlet distributions, a statistically principled way of simulating heterogeneous client data, the advantage ranged from 0.89 to 1.12 percentage points. In a field where fractions of a percentage point can separate competing publications, gains of this size on two independent datasets, against a broad field of rivals, constitute a meaningful demonstration that sensitivity-aware personalization captures something that fixed partitioning schemes miss.

The industrial motivation behind the work is explicit. The authors are affiliated with a school of intelligent manufacturing, and the study was supported by the National Natural Science Foundation of China, the National Key Research and Development Program of China, the China Postdoctoral Science Foundation, and the 111 project, with acknowledgments extending to the Center for Advanced Life Cycle Engineering and the Centre for Advances in Reliability and Safety in Hong Kong. Factories are natural candidates for federated learning because production data is often commercially sensitive and cannot be freely pooled between companies or even between plants of the same company. A method that lets each site tune a shared inspection model to its own conditions, without shipping images anywhere, addresses both the privacy constraint and the heterogeneity problem at once.

Looking at the wider picture, pFedDS adds to a rapidly maturing toolkit for decentralized machine learning, a field whose open problems have been catalogued in a landmark survey by dozens of researchers and whose techniques now span adversarial domain generalization, knowledge distillation, meta-learning, and semi-supervised approaches. The persistent tension at the heart of the enterprise is that collaboration improves models while individuality degrades the shared objective, and every new method is essentially a different answer to how much of each to allow. By making that answer a continuously varying, data-driven property of every single weight, the Jiangnan University team offers one of the more granular answers yet proposed, and their results on real factory-floor data suggest the approach could matter wherever distributed devices must learn together while respecting the stubborn uniqueness of the data they hold.

Subject of Research: Parameter domain sensitivity based personalized federated learning for non-IID data

Article Title: pFedDS: parameter domain sensitivity based personalized federated learning

Article References: Zhao, Z., Su, L., & Li, K. (2026). pFedDS: parameter domain sensitivity based personalized federated learning. Cluster Computing, 29(13), Article 761. https://doi.org/10.1007/s10586-026-06578-9

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06578-9

Keywords: federated learning, personalized federated learning, non-IID data, parameter sensitivity, machine learning, defect detection, industrial inspection, distributed training, data privacy, adaptive learning rates, Cluster Computing, pFedDS

The research was motivated by a practical question facing higher education. Large language models have already woven themselves into student life, with surveys showing that a large majority of university students use generative AI tools beyond formal coursework. Commercial systems such as ChatGPT offer strong performance, but they raise concerns about data privacy, cost, and institutional control. Locally deployable models promise an alternative: a university can run them on its own hardware, keep student conversations off third-party servers, and fix the model version. The catch is that the capabilities of these smaller models, especially in authentic learning interactions rather than tidy one-shot tests, have remained poorly characterized. The team, led by Marcel Völschow of Hamburg University of Applied Sciences together with colleagues at DESY and Helmholtz-Zentrum Dresden-Rossendorf, set out to close that evidence gap.

To measure domain competence, the researchers built mlphys101, a new multiple-choice benchmark of introductory physics questions drawn, with permission, from a publicly available Physics 101 question bank. After removing image-dependent and duplicate items, 731 questions remained. Crucially, the team sorted every question into one of five difficulty categories: replication of definitions, replication of physical facts, conceptual physics and qualitative reasoning, single-step quantitative reasoning, and multi-step quantitative reasoning. This taxonomy matters because it separates what a model merely memorized from what it can actually apply. The questions were translated into German using GPT-4 Turbo and reviewed by a trained physicist and native speaker, and translations into Italian, Polish, and Spanish were also prepared to support future multilingual comparisons.

The model under examination, Gemma 3 27B, was chosen as the best compromise between capability and feasibility for local inference on a 48-gigabyte GPU memory budget. The researchers tested three quantization variants, labeled Q4_K_M, Q6_K, and Q8_0, which compress the model’s numerical weights to different precisions to trade size and speed against quality. Each variant sat for sixteen full exams with different random seeds, served by the llama.cpp inference engine with a 32,768-token context buffer. The prompt instructed the model to think step by step, explain its reasoning in German, and end with a keyword followed by the letter of the correct option. The team deliberately avoided forcing the model into a structured output format, because recent work has shown that format restrictions can impose a substantial accuracy penalty on open-weight models, degrading reasoning even when parsing becomes easier.

The benchmark results were strong. Across all runs and difficulty levels, the model achieved accuracies between 84 and 98 percent. It performed best and most consistently on definitions and single-step quantitative problems, with medians in the mid to high 90 percent range, and dipped to around 90 percent on conceptual questions. Multi-step quantitative reasoning proved the hardest and least consistent category, showing the largest variability across runs. Surprisingly, the quantization level made little difference: higher precision did not yield a uniform improvement, suggesting that even a compressed version of the model retains most of its physics competence. Of more than 34,000 generated responses, the evaluation pipeline failed to extract an answer in only a single case, in which the model concluded that none of the five options was correct.

To put these numbers in context, the researchers also ran the same benchmark against a commercial cloud baseline, OpenAI’s GPT-5.6-Luna. The frontier model achieved consistently near-ceiling accuracy across all five categories, ranging from 0.96 to 0.99, with little variation between repeated runs. The comparison showed that the remaining gap between the local and commercial models was concentrated precisely where tutoring matters most: conceptual understanding and multi-step reasoning. On definitions and straightforward formula application, the compact local model was nearly indistinguishable from its cloud-based rival. But the study’s most important finding came next, when the benchmark champion was asked to do something far messier than answering multiple-choice questions.

In the classroom field experiment, the team deployed Gemma 3 27B locally on a university workstation and exposed it to students through a ChatGPT-like web interface built with Gradio. Thirty-two first-semester electrical and information engineering students took a physics mock exam in which the chatbot, nicknamed Emmy, was their only permitted aid besides a calculator. The system prompt cast the model as an experienced physics teacher who gives hints step by step, reveals full solutions only on request, and remains patient and friendly. Conversations were not saved, protecting student privacy, though students could flag particularly good or bad answers. Afterward, seventeen students completed a detailed questionnaire about their experience.

The survey revealed a sobering picture. Overall helpfulness received a mean rating of only 2.7 out of 5, even though the comprehensibility of explanations scored notably higher at 3.4 and support for understanding fundamental concepts reached 3.5. Fourteen of the seventeen respondents said they had to reformulate their prompts at least once to get a satisfactory answer, and nine had to do so more than once. Fewer than half of the submissions unambiguously agreed that the chatbot understood their questions. Free-text comments pointed to incorrect numerical values and formulas, including one cited error involving the cross-sectional area of a sphere, and to the model losing track of quantities provided earlier in the conversation. Response speed ranked dead last among the system’s features, with fourteen of seventeen students placing it last, reflecting inference delays when many students used the single server simultaneously.

The contrast between the two halves of the study is the paper’s real contribution. On a controlled, single-turn benchmark, the 27-billion-parameter model demonstrated substantial knowledge of introductory physics, approaching commercial performance on many task types. In a genuine multi-turn tutoring setting, that competence did not translate into reliable behavior. Tutoring, the authors argue, demands capabilities that benchmarks never test: interpreting incomplete or conversational student prompts, retaining information across turns, calibrating the level of assistance, and producing explanations that are both physically correct and pedagogically useful. A system that misreads a question, misses a misconception, or offers a plausible but wrong intermediate step may reinforce the very difficulties it is meant to resolve. High benchmark accuracy, in other words, is a necessary but insufficient indicator of tutoring suitability.

The authors are careful about the limits of their work. The field experiment involved a small, single-cohort sample, relied on self-reported perceptions rather than measured learning gains, and deliberately created a worst-case scenario in which students had no other assistance. The benchmark questions were publicly available and could in principle appear in training data, and the reported accuracies measure answer correctness rather than the coherence of the underlying reasoning. Even so, the practical message for universities is clear. Locally deployed language models can genuinely support selected introductory physics activities, offering institutional control over data and infrastructure that commercial services cannot match. But before any model, local or cloud-based, is entrusted with real students, it should be evaluated not only on domain benchmarks but under realistic, multi-turn interaction conditions. Under the conditions studied here, Gemma 3 27B is better described as a potentially useful supportive tool than as a standalone tutor, and the benchmark-to-classroom gap it exposed is one that every educational AI deployment should be tested for.

Subject of Research: Evaluation of a locally deployed large language model for undergraduate physics tutoring using a new benchmark and a classroom field experiment

Article Title: Evaluating Gemma 3 27B for local undergraduate physics tutoring

Article References: Völschow, M., Buczek, P., Riefer, P., Jasko, A., & Steinbach, P. (2026). Evaluating Gemma 3 27B for local undergraduate physics tutoring. Discover Artificial Intelligence, 6(1), Article 1392. https://doi.org/10.1007/s44163-026-02296-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02296-8

Keywords: large language models, Gemma 3 27B, physics education, AI tutoring, benchmark evaluation, mlphys101, quantization, local deployment, classroom field experiment, educational technology, chain-of-thought prompting, data privacy

The new framework, called FL-EF²DP for Federated Learning-based Early Financial Fraud Detection and Prevention, was published in the journal Discover Artificial Intelligence by Mudit Chaturvedi, Shilpa Sharma, and Gulrej Ahmed. Its central premise addresses a long-standing tension in the financial industry: the machine learning models that detect fraud best are trained on enormous volumes of transaction data, yet that data is among the most sensitive information banks hold, governed by stringent privacy regulations that make large-scale sharing impractical and frequently legally prohibitive. Centralized data aggregation, the traditional route to building powerful models, introduces risks of data breaches, single points of failure, and regulatory non-compliance that restrict both real-world deployment and cross-institutional collaboration.

Federated learning offers a way out of this impasse. Under the federated paradigm, a central server initializes a global fraud detection model and distributes its parameters to participating institutions, such as banks. Each institution trains the model locally on its own private transaction records, never transmitting a single raw transaction. Instead, only the updated model weights are sent back to the server, which merges them using the FedAvg averaging algorithm to produce an improved global model. This cycle repeats over multiple rounds until the model converges, at which point the final version is redeployed at every participating institution for real-time inference. The result is a model that has effectively learned from the collective experience of many organizations without any of them ever exposing their customers’ data.

At the heart of the framework sits a convolutional neural network, a deep learning architecture more commonly associated with image recognition. The authors acknowledge that financial transaction data is fundamentally tabular, but they transform the feature vectors into structured two-dimensional matrices before processing. This allows convolutional filters to capture local feature interactions and hierarchical patterns across transaction characteristics that manual feature engineering might miss. The choice of a CNN was also pragmatic: compared with recurrent and attention-based architectures such as LSTM, GRU, or Transformer models, convolutional networks require fewer trainable parameters and less communication overhead, a significant advantage in federated environments where model updates must be shuttled between clients and a server over many rounds. The specific network used in the experiments comprises four convolutional layers and two dense layers, with batch normalization before each pooling operation, a dropout rate of 0.20 to curb overfitting, ReLU activations throughout, a softmax output layer, categorical cross-entropy loss, and the Adamax optimizer.

What distinguishes FL-EF²DP from earlier federated fraud detectors is not a novel optimization technique but the integration of several complementary components into a single operational pipeline. Beyond the federated CNN, the system includes guideline-based control verification, which assesses each institution’s security posture against machine-readable rules drawn from frameworks such as ISO security controls, NIST cybersecurity recommendations, and RBI financial security guidelines. The module collects device- and system-level signals, including firewall status, antivirus operation, access control policies, encryption, and authentication methods, then computes a compliance score and a control severity score reflecting any gaps. An institution that fails to implement multi-factor authentication, real-time transaction monitoring, or periodic access-control reviews sees its compliance score fall and its severity score rise.

These infrastructure assessments are then fused with the behavioral fraud probability produced by the CNN. The risk fusion module computes a consolidated score using the formula Z-risk equals phi times P-F plus beta times U-gap plus gamma times C-risk, where P-F is the fraud probability, U-gap is the normalized guideline violation severity, and C-risk is a contextual risk score derived from transaction-specific and operational characteristics. In the current implementation the weights are fixed at 0.5, 0.3, and 0.2 respectively, summing to one, with a threshold of 0.7 triggering automated mitigation. This design means that a transaction conducted on a vulnerable device can be flagged even when its behavioral fraud likelihood is only moderate, capturing the reality that fraudulent operations are shaped by multiple interacting elements rather than transaction behavior alone.

When the fused risk score crosses the threshold, the system acts immediately. The action and response component, running locally on bank systems to minimize latency, can suspend transactions, require enhanced two-factor authentication, terminate sessions, block clients, or notify a security operations unit. A recommendation report generator then compiles compliance documentation covering system parameters, identified guideline deficiencies with severity classifications, relevant regulatory citations, and actionable remedial steps drawn from a knowledge base. The authors emphasize that this shifts the paradigm from passive detection to proactive prevention, allowing institutions to intervene during the transaction evaluation phase, before further fraudulent transactions, financial losses, or account breaches occur.

The experimental results are striking. Testing was conducted on two widely used benchmarks: the Credit Card Fraud Detection dataset, containing 284,807 real-world European credit card transactions from 2013 with only 492 frauds, a fraud ratio of 0.172 percent, and the PaySim Synthetic Financial Dataset, comprising roughly 6.36 million mobile money transactions with about 8,213 fraud instances, a ratio of approximately 0.13 percent. Both are severely imbalanced, mirroring real-world conditions. The researchers simulated 100 federated clients with a participation ratio of 0.3, meaning 30 clients joined each communication round, and distributed data using a Dirichlet distribution with a parameter of 0.5 to create moderate statistical heterogeneity. On the PaySim dataset, FL-EF²DP achieved 94.12 percent accuracy, 93.88 percent precision, 92.36 percent recall, and an F1-score of 93.11 percent, outperforming competing methods including FED-SPFD at 91.90 percent accuracy and Transformer-LOF-RF at 89.97 percent, while JNBO-SpinalNet fared worst. On the credit card dataset, the full framework reached 96.25 percent accuracy against 91.77 percent for a centralized CNN baseline. An ablation study confirmed that each added component, from federated learning to risk fusion to the complete pipeline, contributed measurable gains, and the framework also achieved the lowest loss values and fastest convergence over 100 epochs.

The authors are candid about the system’s limitations, and these reveal where the field must go next. The current implementation assumes an honest-but-curious server that follows the protocol faithfully but might attempt to infer information from model updates, and it does not yet incorporate formal privacy-enhancing technologies such as differential privacy, cryptographic secure aggregation, homomorphic encryption, or secure multi-party computation. Privacy protections currently stem from data localization and decentralized learning rather than explicit cryptographic guarantees, leaving the system potentially vulnerable to gradient inversion, model reconstruction, and membership inference attacks documented in the federated learning literature. The fixed weighting scheme in the risk fusion module, chosen for simplicity and interpretability, may also benefit from adaptive methods using reinforcement learning or attention mechanisms. Future work will additionally explore graph neural networks and temporal graph learning to capture relationships among users, merchants, devices, and accounts that a CNN cannot represent, along with time-aware evaluation protocols based on chronological transaction sequences.

Even with those caveats, the study offers a compelling glimpse of how financial institutions might collaborate against fraud without surrendering the data their regulators and customers demand they protect. The framework’s computational footprint is modest, with each client storing only its local dataset and a single copy of the CNN, and communication costs reduced by the partial participation scheme. As fraud losses climb toward 40 billion dollars a year and criminals adapt ever faster to static defenses, systems that combine collaborative intelligence with rigorous privacy preservation may prove not just technically elegant but operationally essential for the future of secure digital finance.

Subject of Research: Privacy-preserving federated deep learning for early financial fraud detection and prevention

Article Title: A privacy preserving federated deep learning system for early financial fraud detection and prevention

Article References: Chaturvedi, M., Sharma, S., & Ahmed, G. (2026). A privacy preserving federated deep learning system for early financial fraud detection and prevention. Discover Artificial Intelligence, 6(1), Article 1393. https://doi.org/10.1007/s44163-026-02099-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02099-x

Keywords: federated learning, financial fraud detection, deep learning, convolutional neural networks, data privacy, FedAvg, risk fusion, credit card fraud, PaySim, regulatory compliance, automated mitigation, distributed machine learning

The motivation is rooted in a familiar limitation of critical care monitoring. Conventional ICU devices excel at point measurements—heart rate, blood pressure, oxygen saturation—but they attach to the patient and capture little of the spatial information written on the body’s surface. Patterns of peripheral perfusion, localized inflammation, or uneven temperature distribution can carry early signals of deterioration, yet standard multiparameter monitors are largely blind to them. Infrared thermography offers a way in: a radiometric thermal camera can map surface temperature remotely, without touching the patient, and previous studies have used thermal imaging to estimate respiratory rate and other vital signs. The catch, the authors note, is that most of that evidence comes from controlled laboratory prototypes, small cohorts, or algorithm-centered studies that never faced the realities of a hospital ward.

Those realities are formidable. Any system deployed in a modern hospital must pass through institutional information technology governance: network segmentation, cybersecurity policy, bandwidth limits, authentication requirements, and data protection law. Public cloud architectures, common in commercial Internet of Medical Things products, are often difficult or impossible to approve in European hospitals that must comply with the General Data Protection Regulation and maintain local control over patient-related data. The Madrid team therefore built their system entirely on-premise. Seven acquisition nodes—each a Raspberry Pi 4B single-board computer paired with a low-cost radiometric thermal camera of 256 by 192 pixel resolution—were installed at ICU beds, with the architecture designed to scale to all thirteen beds in the unit. Each node captured one thermal image per second, tagged it with timestamps and temperature statistics, and streamed it over the hospital’s existing Wi-Fi network to a local server for storage, indexing, and near real-time visualization. No parallel network was created, and no external cloud service touched the data.

Privacy was engineered into the hardware rather than bolted on afterward. Unlike clinical video systems, which typically require post-hoc de-identification steps such as blurring faces before footage can be released, the low-resolution radiometric thermal sensor cannot resolve the optical facial features needed to identify a person in the first place. Non-identifiability is a structural property of the acquisition modality itself, not a downstream processing step that could be omitted, misconfigured, or reversed. The database stored only bed identifiers and technical metadata, image acquisition was conditional on informed consent and activation by authorized personnel, and the study was approved by the relevant research ethics committee under approval number CEIm 23/63, conducted in accordance with the Declaration of Helsinki.

The first deployment phase delivered a humbling lesson in systems engineering. In its initial configuration, with synchronous image transmission and accumulating queues, the system buckled under sustained load. Analysis of 127,618 latency samples revealed a mean end-to-end delay—the time between image capture at the bedside and availability on the dashboard—of 26,344 milliseconds, with a median of 20,616 milliseconds and a 95th percentile reaching nearly 84 seconds. Images were piling up in transmission queues faster than they could be processed, turning a nominally real-time system into a delayed archive. The fix came not from replacing hardware but from rethinking the data flow: the team decoupled image reception from persistence, prioritized the most recent frame for visualization, improved queue management, and adjusted capture and transmission parameters, allowing older frames to be retransmitted asynchronously when bandwidth allowed.

The results of that architectural overhaul were dramatic. Across 6,917,464 subsequent latency samples, mean end-to-end latency fell to 126 milliseconds—a 99.52 percent reduction—with a median of 117 milliseconds and a 95th percentile of just 207 milliseconds. Crucially, the improvement extended to the upper tail of the distribution, confirming that prolonged queue accumulation had been eliminated rather than merely reduced. The system had been transformed, in the authors’ framing, from a delayed acquisition platform into a genuinely near real-time monitoring infrastructure, achieved purely through software and data-flow design on the same inexpensive sensing hardware.

With the optimized architecture in place, the team ran a structured ten-day evaluation with all seven nodes active. The system generated 5,763,312 thermal images, roughly 1.47 terabytes of data at an average of 275 kilobytes per image and about 147 gigabytes per day. The aggregate acquisition rate of 6.67 images per second came remarkably close to the nominal target of seven, indicating that the system held nearly full capture capacity under real operating conditions. Latency remained stable throughout the evaluation, showing no progressive deterioration as the stored dataset grew into the millions of images—a sign that the hybrid persistence strategy, with raw images in a hierarchical file system and metadata indexed in a PostgreSQL database, could sustain continuous long-term operation. Server resource use stayed almost trivially low, with mean CPU utilization of just 0.66 percent and mean RAM usage of 6.93 percent, leaving ample headroom for the planned expansion to thirteen beds.

The system was not flawless, and the failure modes it exhibited are themselves informative. Effective availability during the evaluation was 95.29 percent, with the shortfall traced almost entirely to transient wireless connectivity interruptions and local node issues rather than any failure of the central server or storage infrastructure. Telemetry recorded 44 Wi-Fi interface drop or reconnection events, cumulatively lasting 106,218 seconds, with individual incidents ranging from the ten-second minimum observable interval to a maximum of 40,807 seconds. Some nodes, labeled BOX7 and BOX8, showed larger temporal discontinuities, while BOX9 and BOX11 behaved most consistently. The team also contended with clock synchronization drift between nodes and server, addressed through network time synchronization and a dual timestamp strategy, and with power supply instability in some devices, mitigated by stabilizing the supply. The pattern is a familiar one in hospital deployments: the bottleneck is rarely computation or storage, but the shared, policy-constrained wireless environment that research systems must coexist with.

The authors are careful to frame this as foundational infrastructure rather than a clinical tool. The dashboard at this stage provides no alarms, no diagnostic interpretation, and no automated decision support, and the reported temperature metrics were computed over the entire thermal frame rather than a patient-isolated region, so they cannot be read as validated measurements of body temperature—bedding and nearby equipment could fall within the field of view. This deployment is explicitly the first of two phases in the approved research protocol, establishing technical feasibility, reliability, and governance compliance before a planned second phase evaluates clinical utility, including automated detection of patient falls, self-extubation, pressure ulcer development, and cardiac arrest, a role in differentiating septic, cardiogenic, and hypovolemic shock, and monitoring of postoperative recovery of normothermia. The team also acknowledges honest limitations: transport between nodes and server used HTTP without TLS encryption, a risk mitigated by network-layer isolation but flagged as a security-hardening priority; no formal penetration testing was performed; the study was single-center with one hardware configuration; and scalability to the full thirteen-bed configuration was projected from operational trends rather than directly tested. The system remains a pre-product research infrastructure, not a certified medical device.

Even with those caveats, the study offers a rare and valuable data point: a complete, quantified account of what it actually takes to run a distributed medical sensing network inside a live ICU. The practical lessons the authors distill—deploy progressively rather than all at once, coordinate closely with hospital IT teams from day one, build local buffering and asynchronous transmission into the architecture for resilience, and evaluate operational metrics like latency, availability, and storage growth rather than sensor accuracy alone—read as a reference model for anyone attempting similar deployments. As hospitals edge toward AI-assisted, contactless monitoring, the hardest problems may not be in the algorithms at all, but in the unglamorous plumbing of queues, clocks, tokens, and Wi-Fi. In Madrid, at least, that plumbing now works, one thermal frame at a time.

Subject of Research: On-premise thermal Internet of Medical Things monitoring in an intensive care unit

Article Title: Real-World Deployment and Operational Evaluation of an On-Premise Thermal IoMT System in an Intensive Care Unit

Article References: Avilés, E., Lafuente, J.-L., González, S., & Beunza, J.-J. (2026). Real-World Deployment and Operational Evaluation of an On-Premise Thermal IoMT System in an Intensive Care Unit. Journal of Medical Systems, 50(1), Article 144. https://doi.org/10.1007/s10916-026-02455-5

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02455-5

Keywords: Internet of Medical Things, thermal imaging, intensive care unit, patient monitoring, on-premise architecture, hospital IT governance, Raspberry Pi, infrared thermography, latency optimization, data privacy, edge computing, critical care

The audit was deliberately designed to measure what universities show the world, not what they may do behind closed doors. Abdulrahman and Ahmed argue that public visibility is itself a governance issue: students, lecturers, parents and policymakers rely on official websites to understand what counts as acceptable behaviour, and guidance that is hidden, outdated or written only in technical English cannot shape everyday academic practice. Between 12 and 20 June 2026, the researchers applied an identical five-step search protocol to each of the 30 institutions, drawn purposively from the Kurdistan Regional Government’s public list of 38 higher-education entities. They searched for terms ranging from plagiarism and academic integrity to ChatGPT, generative AI, privacy policy and data protection, checking Kurdish and Arabic sections of websites where relevant. Crucially, the team treated absence of evidence carefully: a score of zero meant only that nothing was publicly visible during the search period, not that a university had no internal policy at all.

Each university was assessed across 11 dimensions, spanning plagiarism policy, academic integrity guidance, generative AI guidance, acceptable and unacceptable AI use, AI-use disclosure, assessment redesign, student AI literacy, staff guidance, privacy and data protection, institutional governance, and Kurdish or local-language accessibility. Every dimension was scored from 0 to 3, giving a maximum institutional score of 33. The researchers specified the framework before scoring, refined it through review by six independent experts in academic integrity, artificial intelligence and higher-education policy, and then tested it for consistency. A subset of 10 universities was independently double-coded, covering 110 coding decisions: the two coders agreed exactly on 105 of them, a 95.5 percent agreement rate, with a quadratic weighted kappa of 0.975, an exceptionally high figure. A sensitivity analysis shifting the classification thresholds by one or two points confirmed that the overall pattern was robust.

The headline finding is the distribution of institutions across a proposed four-stage readiness model that moves from plagiarism control, through academic integrity and AI awareness, to full AI governance. Twenty-one of the 30 universities, 70 percent, sat at Stage 1, indicating either low public visibility or a narrow orientation toward plagiarism control. Five institutions, 16.7 percent, reached Stage 2, reflecting broader academic-integrity provisions; three, 10 percent, showed genuine AI awareness at Stage 3; and only one university, 3.3 percent, demonstrated a more developed public-facing AI governance position at Stage 4. Public and private institutions performed almost identically, with mean scores of 4.63 and 4.43 out of 33 respectively, though the single Stage 4 case was a private university. The researchers are careful to note that the sample was purposive rather than probabilistic, so no claims about sector-wide population effects are made.

Direct evidence of generative-AI-specific policy was strikingly rare. Only two universities, 6.7 percent of the sample, showed direct public evidence of GenAI policy or guidance, four more showed partial evidence, and the remaining 24, fully 80 percent, had no clear public GenAI policy visible during the search window. This is perhaps the study’s most consequential finding: while students and lecturers are already using AI tools daily for translation, grammar support, coding assistance and summarisation, the overwhelming majority of institutions offer no visible guidance on what is permitted, what must be disclosed, or what data should never be pasted into a public chatbot. The strongest performers were Tishk International University, followed by the American University of Iraq–Sulaimani, the University of Kurdistan Hewlêr, the American University of Kurdistan, Koya University and the University of Duhok, yet even these combined institutions did not show complete evidence across all 11 dimensions.

The pattern of strongest and weakest dimensions tells a coherent story about where institutional attention has been directed. Plagiarism policy scored highest at a mean of 0.73 out of 3, followed by academic integrity at 0.67 and general AI awareness at 0.60. At the bottom of the table sat the dimensions that matter most for responsible AI use: AI-use disclosure averaged just 0.13, acceptable-use guidance 0.17, privacy and data protection 0.20, and AI-aware assessment redesign 0.23. The authors argue these gaps are not minor administrative details but the core of meaningful governance. Without disclosure rules, students cannot know how to report AI assistance; without acceptable-use examples, they may assume all AI use is either forbidden or freely allowed, both of which corrode fairness; without privacy guidance, sensitive student data and unpublished research can flow into commercial AI systems unchecked; and without assessment redesign, universities default to unreliable detection software and punishment.

Underlying the whole study is a conceptual argument: generative AI has broken the traditional definition of plagiarism. AI-assisted work is rarely simple copying from an identifiable source. A student may use a chatbot to improve grammar, translate between Kurdish and English, generate an outline, explain a concept, summarise readings or draft code. Some of these uses support learning; others may quietly replace the student’s own intellectual contribution. Plagiarism policies remain necessary, the authors contend, but they are no longer sufficient. The study points to international practice, including the Artificial Intelligence Assessment Scale, which helps educators communicate different permitted levels of AI use depending on the learning outcome, shifting the conversation from detection toward transparent assessment design. AI-aware assessment options such as oral defences, process logs, reflective writing, drafts with feedback histories and tasks requiring students to justify their reasoning offer a sturdier defence than similarity checkers, which are notoriously unreliable when AI-generated text has been human-edited.

The multilingual context of Kurdistan higher education adds a distinctive layer to the problem. Students and lecturers routinely work across Kurdish, Arabic and English, and AI tools are heavily used for translation and English academic-writing support. The study insists that AI governance in the region must be locally grounded: guidance should be available in Kurdish and, where relevant, Arabic and English, with examples drawn from regional realities such as multilingual writing, translation and local case studies. Policies written only in technical English, the authors warn, will not reach all students equally, making language accessibility a matter of equity rather than mere convenience. Recent research on Kurdish English-language teachers, showing that digital literacy, technophilia and technophobia all shape how AI is integrated into teaching, reinforces the case for institutional literacy programmes rather than purely punitive responses.

From the findings, the researchers derive a practical toolkit: the four-stage readiness model and an 11-component AI governance checklist that universities, quality-assurance units and policymakers can use to audit themselves. Their recommendations are concrete. Universities should update plagiarism and integrity policies with GenAI-specific examples, create a dedicated and easy-to-find public AI guidance page, provide disclosure templates and assignment-level rules, support lecturers in redesigning assessments, publish AI-literacy resources covering hallucination, fabricated references, bias and over-reliance, and specify what information must never be entered into public AI tools. Responsibility for AI governance should be assigned to a named body, whether a quality-assurance unit, digital-learning team or cross-institutional working group, and reviewed regularly, because the technology is changing far faster than policy cycles.

The authors acknowledge the limits of their method. A website audit measures visibility, not implementation; universities may have internal rules that never surfaced in the search. Websites are dynamic, and the team has published a supplementary evidence workbook and audit archive on Zenodo preserving row-level scores, exact search queries, dates and negative-search logs to make the study reproducible. Equal weighting of the 11 dimensions is a simplification, and document analysis cannot capture the depth of classroom practice. Yet the central conclusion stands with unusual clarity for a policy audit: plagiarism rules are necessary but no longer sufficient. Generative AI has transformed academic writing, authorship, assessment and digital responsibility, and universities that communicate nothing publicly about it leave students to guess where support ends and misconduct begins. The Kurdistan Region’s universities, this study suggests, now have a roadmap for the journey from plagiarism control to genuine, transparent and multilingual AI governance; the question is how quickly they will travel it.

Subject of Research: Public-facing generative AI governance readiness in Kurdistan Region higher education institutions

Article Title: Mapping generative AI governance readiness in Kurdistan Region universities

Article References: Abdulrahman, A. A., & Ahmed, S. K. (2026). Mapping generative AI governance readiness in Kurdistan Region universities. Discover Education, 5(1), Article 1132. https://doi.org/10.1007/s44217-026-02242-x

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02242-x

Keywords: generative AI, higher education, academic integrity, plagiarism, AI governance, ChatGPT, assessment redesign, AI literacy, data privacy, Kurdistan Region, policy audit, digital learning

The research team conducted the review according to the Joanna Briggs Institute methodology for scoping reviews and reported it following the PRISMA-ScR guidelines, a rigorous framework designed to make evidence syntheses transparent and reproducible. They searched the Web of Science and Ovid databases, including Ovid MEDLINE and APA PsycInfo, for literature published between January 2010 and February 2025. The initial search retrieved 1,173 articles, which the team screened down to 39 studies that underwent full data analysis through narrative synthesis. That funnel shape — from nearly twelve hundred records to fewer than forty included studies — reflects both the growing but still fragmented nature of this research field and the strict inclusion criteria the reviewers applied.

The scope of the review is deliberately broad. Rather than focusing on a single disease, the authors set out to synthesize the needs and experiences that parents of children with chronic conditions express across different diagnosis categories. This cross-diagnostic approach is what gives the findings their punch: recurring themes emerged regardless of whether the child had autism, diabetes, asthma, congenital heart disease, or obesity. In other words, the review suggests that the digital needs of caregiving parents are not primarily disease-specific but reflect the shared architecture of chronic pediatric care itself — information, coordination, communication, and emotional support.

Technically, the e-Health landscape covered by the included studies spans several distinct modalities. Mobile applications delivered on smartphones, web-based platforms accessible through browsers, and video conferencing tools connecting families with clinicians all featured in the literature. Parents’ experiences with these technologies were generally positive, particularly where the tools improved day-to-day monitoring of the child’s condition, strengthened parental empowerment, and expanded access to care. For families living far from specialist centers, or juggling work and caregiving, the ability to transmit data, receive feedback, and consult professionals remotely can transform the logistics of managing a chronic illness.

Yet the review is equally clear about what parents want these tools to be. Across conditions, parents consistently emphasized the need for accessible and user-friendly interfaces — a deceptively simple requirement that many digital health products fail to meet. They also wanted reliable and comprehensible health information, meaning content that is trustworthy, up to date, and written in language a stressed parent can absorb at midnight. A third cluster of needs centered on care coordination and communication: tools that help families organize appointments, track medications, share information with multiple providers, and communicate efficiently with healthcare professionals. In fragmented health systems where a chronically ill child may see pediatricians, subspecialists, therapists, and school nurses, the parents often become the de facto coordinators, and they are asking technology to share that load.

Beyond logistics, the review highlights a more human dimension. Parents valued social support and psychological resources delivered through e-Health applications — features that connect them with other families facing similar challenges or provide coping strategies for the emotional strain of long-term caregiving. Chronic pediatric illness is not only a medical problem; it reshapes family life, finances, and mental health. The finding that parents actively seek psychological and peer support from digital tools underscores that well-designed e-Health is not merely a data pipe between home and hospital but a potential lifeline against isolation.

At the same time, the enthusiasm is tempered by persistent concerns. Data privacy remained a recurring worry among parents, an issue that has only grown in salience as health applications collect increasingly granular information about children. Parents also reported practical barriers: technical difficulties that stall adoption, limited customization that prevents tools from fitting their child’s specific situation, and a fear that increased digitalization could erode the personal contact with providers that many families treasure. That last concern is particularly important for developers and policymakers, because it reframes e-Health not as a replacement for clinical relationships but as an adjunct whose value depends on preserving human connection.

One of the most consequential findings of the review concerns the trajectory of care. The authors found that evidence on parental needs along the care pathway was limited, with early phases after diagnosis and long-term management phases better represented in the literature than transitional periods — for example, the handover from pediatric to adult healthcare services that adolescents with chronic conditions must eventually navigate. Needs and experiences were frequently repeated across different modalities and conditions, which the authors interpret as pointing to common priorities for digital health design. But the dynamic evolution of those needs over time remains under-studied, leaving a gap for future research on how a parent’s requirements change as a child grows, gains autonomy, and eventually takes over their own disease management.

The implications for the digital health industry are straightforward. If effective tools for this population should integrate accessible design, reliable information, coordination features, social support, and robust data protection — as the review concludes — then product development in pediatric e-Health should start from these cross-cutting parental priorities rather than from diagnosis-specific feature lists. The authors argue that parents of children with chronic conditions articulate broad, cross-cutting needs that go beyond diagnosis-specific management, a conclusion that could inform everything from user interface decisions to regulatory frameworks for pediatric health data.

For the growing population of families affected by childhood chronic illness, the message of this review is one of cautious optimism. The technology to support them largely exists, and where it has been deployed well — in mobile apps, web platforms, and video consultations — parents report genuine improvements in monitoring, empowerment, and access to care. What remains is the harder work of alignment: designing systems that are simple enough for an exhausted parent at the end of a long day, transparent enough to earn their trust with their child’s data, and flexible enough to evolve with a family’s journey from diagnosis through transition to adult care. The 39 studies synthesized here provide the clearest map yet of what that alignment must achieve.

Subject of Research: Parental needs and experiences with e-Health applications in chronic pediatric care

Article Title: Parental needs and experiences with e-Health applications in chronic pediatric care – a scoping review

Article References: Grüter, H., Loeffler, A., Stähler, L., Gönen, I., Prüfe, J., De Bock, F., & Pischke, C. R. (2026). Parental needs and experiences with e-Health applications in chronic pediatric care – a scoping review. BMC Pediatrics, 26(1), Article 910. https://doi.org/10.1186/s12887-026-07737-y

Image Credits: AI Generated

DOI: 10.1186/s12887-026-07737-y

Keywords: e-Health, pediatrics, chronic conditions, scoping review, parental needs, digital health, care coordination, data privacy, mobile health apps, telemedicine, health informatics, caregiver support

The researchers began with an enormous haystack. A systematic search of the Web of Science database initially returned more than 36,000 studies touching on artificial intelligence and higher education. Through successive filtering by topic, language, publication type, and additional keywords covering management, teaching, learning, research, and engagement, the pool narrowed to 1,998 candidates. Applying strict inclusion criteria—empirical studies describing actual cases of AI integration at system level rather than isolated chatbot experiments in a single course—the team settled on 49 case studies drawn from 25 countries, including China, the United States, Germany, Ecuador, South Africa, and Ukraine.

Using thematic analysis following the classic qualitative sequence of data reduction, data display, and conclusion drawing and verification, the authors coded the empirical sections of each study. In a striking methodological twist, they also uploaded the selected papers to NotebookLM, a generative AI platform, and asked it to propose its own codes and themes. The researchers reported no visible difference between their human-derived codes and the AI’s suggestions, though they adopted several AI-proposed technical codes covering system architectures, layered ICT infrastructure, programming languages, parameter optimisation, 5G edge computing, and cloud servers.

The first major finding concerns purpose. Universities, the analysis shows, deploy AI across four broad domains: teaching, research, external engagement, and administration. In teaching, AI powers personalised learning pathways, intelligent tutoring systems, automated assessment, and adaptive platforms that respond to individual cognitive patterns. In research, large language models assist with literature reviews, experimental design, data analysis, and manuscript preparation—one reviewed study claimed AI could compress the time to complete an academic paper from several months to a few weeks. In engagement, institutions use generative tools for marketing content, press releases, admissions inquiries, and alumni relations. In administration, AI algorithms optimise the allocation of human, material, and financial resources, streamline enrolment and financial aid services, and even monitor campus security and waste management.

Yet the benefits come bundled with formidable challenges. The reviewed literature documents difficulties with adoption and implementation, which demand substantial investment in technical expertise, infrastructure, and specialised software, alongside sustained coordination between academic and administrative staff. Legacy administrative systems, refined over decades, resist integration with emerging technologies. Data availability is another bottleneck: AI models require large, high-quality datasets, but underdeveloped collection infrastructure, privacy regulations, and small enrolments can cripple predictive accuracy. Add ethical concerns about privacy and algorithmic bias, plus human resistance from staff worried about job security and deprofessionalisation, and the picture becomes considerably more complicated than vendor brochures suggest.

The heart of the study is its eight-step framework for technical and institutional integration. Step one is AI system design, where experts define the university ecosystem—often through an ontology of classes describing the academic environment—and choose a layered, multi-platform architecture spanning data collection, analysis, and application. Step two is algorithm development, where institutions choose between open-source and paid AI services, prompt-based tools such as ChatGPT or task-oriented systems, and decide whether to build bespoke models, from graph neural networks that optimise educational pathways to hybrid BERT-LDA semantic topic modelling for research evaluation. Machine learning foundations—support vector machines, neural networks, and unsupervised approaches using natural language processing, understanding, and generation—remain the critical root of success.

Step three is data management and preprocessing, which the authors summarise with the mantra ‘Do Data First.’ Real-world university data is messy, drawn from sources as varied as student activity logs, real-time sensor feeds, health data from wearable bracelets, and institutional audit corpora. Cleaning, validating, normalising, and structuring that data—removing duplicates, handling missing values, ensuring interoperability—determines whether downstream AI delivers insight or noise. Step four is system integration, where universities define multi-role designs clarifying how human actors and AI tools interact, then embed AI into learning management systems, student affairs platforms, mental health counselling services, payroll and attendance systems, and communication channels including external APIs and social networks.

The institutional half of the roadmap begins at step five with strategic management policies. Leadership and vision emerge as decisive factors, alongside honest evaluation of the university’s readiness for digital transformation. Policies must prioritise credibility, security, and privacy while aligning with binding external regulation—the EU AI Act, GDPR, UNESCO’s Recommendation on the Ethics of Artificial Intelligence, and national frameworks from Australia to Türkiye. Step six assigns operational management roles, negotiating the division of labour between humans and machines so that AI augments rather than replaces professional judgement. Step seven embeds AI in curriculum and pedagogical design, both by using AI to track rapidly shifting industry demands and by teaching AI literacy itself. Step eight closes the loop with evaluation and performance monitoring, from automatic essay assessment with explanatory feedback to learning analytics dashboards that flag students at risk and generate statistical, vulnerability, audit, and health reports for managers.

Ethics, the authors insist, is not an afterthought but a fifth pillar organised around five themes: data ethics, algorithmic ethics, human-AI interaction and social impact, academic integrity and responsible use, and institutional and managerial ethics. Machine learning trained on small or incomplete datasets produces biased results; algorithms can quietly amplify disparities of gender, race, and socioeconomic status. Overreliance on AI risks eroding critical thinking, and hallucinations compound the danger. The study stresses that undisclosed AI use in assignments or research is comparable to plagiarism, and that transparency about AI’s involvement is essential for reproducibility. Countermeasures range from technical designs such as a ‘Trustworthy Ethical Firewall’ to continuous ethics committee review and institution-wide AI literacy training for students, academics, and administrators alike.

Perhaps the most encouraging message is that AI integration is not the exclusive privilege of wealthy institutions. While a few high-profile cases such as Pennsylvania State University and Purdue University appear among the selected studies, the framework also documents successful deployments in resource-constrained settings: an AI chatbot answering admissions inquiries, an AI-powered mentor helping students comprehend scientific texts, a Moodle-based personalised feedback system, and a smart sketchpad supporting film character design. Framed through socio-technical systems theory, the authors conclude that AI adoption is a strategic initiative rather than a mere technological upgrade—one requiring purposeful leadership, sequenced technical execution, and an ethical culture that evolves beyond compliance into genuine institutional habit. For universities still wondering where to start, the answer, apparently, is step one.

Subject of Research: Integration of artificial intelligence systems into university teaching, research, and administrative practices

Article Title: Meta-synthesis of global cases for the adaption stages of artificial intelligence systems into university practices

Article References: Uslu, B., & Leišytė, L. (2026). Meta-synthesis of global cases for the adaption stages of artificial intelligence systems into university practices. Discover Education, 5(1), Article 1109. https://doi.org/10.1007/s44217-026-02169-3

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02169-3

Keywords: artificial intelligence, higher education, meta-synthesis, university management, AI ethics, machine learning, learning analytics, academic integrity, data privacy, institutional policy, AI literacy, educational technology

The headline figure of 86 percent non-compliance conceals a range of distinct violations, each of which carries its own legal weight under European and British data protection law. Two-thirds of the audited sites, 67 percent, began collecting personally identifiable information before users had given any consent at all, transmitting unique user identifiers to third-party analytics and marketing platforms. Under the GDPR and the UK GDPR, consent must be obtained before non-essential cookies are placed on a user’s device, so this practice constitutes a direct infringement rather than a technicality. In effect, the most commercially valuable data flows were already underway while the consent banner was still sitting on the screen, unread and unclicked.

Denial of choice was another widespread problem. Almost a quarter of the sites, 24 percent, offered users no way to refuse tracking whatsoever, and 2 percent displayed no consent banner at all, despite the legal requirement to do so. Even where a reject option technically existed, exercising it could be an ordeal. Only 29 percent of banners allowed users to reject tracking as easily as they could accept it, and on some sites it took as many as fifteen clicks to refuse consent, compared with the single click required to accept. The asymmetry is not subtle. A user who wants to protect their privacy must navigate layered menus and buried settings, while a user willing to be tracked resolves the interaction instantly.

The study catalogued these manipulative layouts under the umbrella term dark patterns, a phrase coined in user-experience research to describe interface designs that steer people toward choices they would not otherwise make. The audit found most banners employed at least one such technique. Visual emphasis of the accept button appeared on 60 percent of sites, making the compliant option the most eye-catching element on screen. Hiding the reject option behind a second layer of navigation occurred on 47 percent of sites, and 29 percent pre-selected privacy-unfriendly settings, requiring users to actively untick boxes to protect themselves. Each of these designs exploits well-documented quirks of human cognition, including the tendency to click prominent, brightly coloured buttons and to accept default options rather than modify them.

To test whether these designs actually change behaviour, rather than merely correlating with poor privacy outcomes, the researchers conducted a controlled online experiment. A sample of 615 UK online gamblers were presented with a simulated betting website carrying one of six different consent banners. The design most commonly used across the gambling industry made participants three to four times more likely to accept tracking than a neutral, one-click alternative. This is a striking effect size for a single interface element, and it demonstrates that the industry’s dominant banner style is not a passive or accidental feature. It functions, in the researchers’ framing, as an active nudge toward the least private option available.

Perhaps the most revealing result concerned how participants felt about their own decisions. Those who accepted tracking rated their choice as a far poorer reflection of their real privacy preferences than those who rejected it, scoring the alignment at 4.4 out of 10 compared with 7.9. In other words, people who accepted tracking under the industry-standard design knew, at some level, that the decision did not represent what they actually wanted. The design, rather than the user, was driving the outcome. Notably, the effect was the same regardless of participants’ level of gambling risk, meaning that both recreational gamblers and those at elevated risk of harm were equally susceptible to the manipulative layouts.

Under GDPR standards, consent is only valid when it is freely given, specific, informed and unambiguous, and it must be as easy to withdraw as it was to give. A consent obtained through a design that makes refusal disproportionately difficult arguably fails the freely given test, which is why the researchers characterise the overwhelming majority of audited sites as non-compliant rather than merely unfriendly. Regulators including the UK Information Commissioner’s Office have previously warned against consent banners that nudge users toward acceptance, but this study provides the first systematic, industry-wide quantification of the problem within gambling, a sector whose business model depends heavily on data-driven customer acquisition and retention.

The stakes extend well beyond data protection compliance. The researchers argue that the data harvested through these designs underpins the personalised marketing and cross-web tracking of gambling customers, and that the characteristics used to identify commercially valuable players overlap substantially with the behavioural markers of gambling harm. In practical terms, the same signals that flag someone as a profitable target, such as frequent sessions, rapid betting patterns and responsiveness to promotions, are also the signals associated with problematic play. Professor Simon Dymond, who supervised the research alongside Dr Martyn Quigley and Dr Jamie Torrance, noted that the behavioural data being harvested is not neutral, and that getting consent right is not just a data protection issue for the industry but an important part of preventing gambling harm.

Jack McGarrigle, the PhD student who led the study, described the audit’s findings as stark, observing that most sites are not giving customers a fair choice about tracking, with some making it a single click to accept and up to fifteen to refuse. He added that the follow-up experiment showed the effect is not incidental, and that the banners work exactly as one would expect, nudging people toward decisions they do not actually agree with. The transparency of the research itself is notable: the data, pre-registration materials and screenshots of all 624 audited sites have been made publicly available through the Open Science Framework, allowing other researchers, regulators and journalists to verify the findings independently.

The study arrives at a moment of intensifying scrutiny of both the gambling industry and the digital advertising ecosystem that sustains it. As regulators weigh how to enforce consent rules against operators whose interfaces are built to circumvent them, the Swansea findings supply a detailed evidentiary baseline: a complete census of a regulated national market, a taxonomy of the dark patterns in use, and experimental proof that those patterns shift real decisions. Whether the response takes the form of enforcement action, mandatory interface standards, or redesign of the consent framework itself, the research makes clear that the current system, in which nine in ten licensed operators fail to offer a fair choice, is not a marginal compliance gap but a structural feature of how the online gambling industry acquires its data.

Subject of Research: GDPR compliance and dark-pattern cookie consent design on UK online gambling websites

Article Title: Nine in 10 UK gambling websites breaching data privacy law, study finds

Article References: Nine in 10 UK gambling websites breaching data privacy law, study finds. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: GDPR, online gambling, cookie consent, dark patterns, data privacy, Swansea University, Gambling Commission, behavioural experiment, gambling harm, digital tracking, consent banners, UK regulation

The core problem the researchers address is one that has haunted privacy-preserving machine learning for years. Techniques such as fully homomorphic encryption, secure multi-party computation, and differential privacy can, in principle, guarantee that data remains usable but not visible. Homomorphic encryption, for example, allows mathematical operations to be performed directly on encrypted values, so the underlying information is never decrypted during computation. Secure multi-party computation lets several parties jointly compute a function over their combined inputs while keeping those inputs hidden from one another. Differential privacy adds carefully calibrated noise so that the contribution of any single individual cannot be reverse-engineered from the final model. In theory, these tools deliver exactly what privacy-conscious institutions need. In practice, they have struggled to keep up with the demands of modern deep learning.

The difficulty lies in overhead. Fully homomorphic encryption can slow training by orders of magnitude, turning what would take hours on unencrypted data into computations that stretch across days or weeks. Secure multi-party computation imposes heavy communication burdens, requiring participants to exchange large volumes of messages for every step of the training process. Meanwhile, the deep architectures that dominate contemporary artificial intelligence, such as ResNet convolutional networks and Transformer models, involve millions or billions of parameters whose gradients must be computed, encrypted, and aggregated at every iteration. The result is a gap between what privacy theory promises and what deployment reality allows, a gap that has kept many hospitals, banks, and agencies from adopting encrypted training at all.

Bao and Su’s framework attacks this gap from four directions at once. The first is a distributed key collaboration mechanism built on Shamir secret sharing, a classical cryptographic technique in which a secret, here an encryption key, is split into multiple shares distributed among different parties. No single participant holds the complete key; only a sufficient quorum of shares can reconstruct it. This design eliminates centralized trust dependence, meaning there is no single server or administrator whose compromise would expose the entire system. If one party is breached or behaves maliciously, the key remains safe as long as the attacker has not collected enough shares, a property that dramatically raises the bar for any would-be adversary.

The second innovation is the heart of the framework’s efficiency gains: the integration of partial homomorphic encryption with Top-k gradient sparsity performed directly in the ciphertext domain. Unlike fully homomorphic encryption, which supports arbitrary computation on encrypted data at enormous cost, partial homomorphic encryption supports a limited set of operations, such as addition, at a fraction of the computational price. Gradient sparsification exploits the observation that in most training iterations, only a small fraction of a model’s gradient components carry meaningful magnitude. By selecting and transmitting only the top-k largest gradient values, the framework slashes the amount of data that must be encrypted and exchanged. Crucially, the researchers perform this selection and aggregation on encrypted values, so the sparsification itself never exposes which parameters matter most, a detail that could otherwise leak information about the training data.

The third component is a lightweight adaptive scheduling strategy that dynamically balances three competing demands: the heterogeneity of participating devices, the desired strength of security, and the overall efficiency of training. In real-world collaborative settings, participants range from powerful data-center servers to modest edge devices, and a rigid protocol that treats them all identically will be throttled by its weakest member. The adaptive scheduler adjusts workloads and security parameters on the fly, allowing faster devices to contribute more while maintaining the cryptographic guarantees that make the whole exercise worthwhile. This kind of systems-level thinking, the authors argue, is what separates a laboratory demonstration from a deployable platform.

Finally, the team wrapped these mechanisms into a unified, end-to-end software system that provides closed-loop protection across the entire machine learning lifecycle, from the moment data enters the pipeline, through encrypted collaborative training, to the point where the finished model is served to end users. This holistic architecture matters because privacy failures often occur not during training itself but at the boundaries, when data is ingested, when intermediate results are exchanged, or when model predictions are exposed. By covering the full pipeline, the system reduces the attack surface that piecemeal solutions leave open.

The experimental results are striking. On benchmark datasets including CIFAR-10, FEMNIST, and a real-world medical dataset, the framework trained ResNet-18 and Transformer models 3.2 times faster than CryptoNets and 2.7 times faster than HE-Transformer, two well-known homomorphic encryption baselines. Communication overhead, often the hidden killer in distributed encrypted training, came in at 42.3 megabytes, dramatically lower than the homomorphic baseline schemes. Perhaps most importantly for practitioners, the accuracy penalty relative to an unencrypted federated averaging baseline was held to just 1.2 percent, a margin small enough that most applications could absorb it without noticing. The authors also report strong scalability and robustness against real-world privacy attacks, suggesting the protections hold up under adversarial pressure rather than only in idealized conditions.

Why does this matter beyond the benchmark suite? Consider a consortium of hospitals that wants to train a diagnostic model on millions of patient records scattered across institutions. Legal frameworks such as data protection regulations often prohibit sharing raw records, and even anonymized data has been re-identified in past studies. Encrypted collaborative training offers a way out: each hospital keeps its records local, gradients are encrypted before leaving the building, and the aggregated model emerges without any participant ever seeing another’s data. Similar scenarios apply to banks pooling fraud-detection intelligence, telecommunications operators improving network models, and government agencies collaborating across jurisdictions. A framework that makes such training fast enough and cheap enough to run on realistic hardware could convert privacy-preserving machine learning from a theoretical curiosity into standard practice.

The study, published in the special issue on Cognitive based Information Processing and Applications 2024, was supported by research projects in Jiangsu higher education institutions, and the authors state they have no competing interests. Its significance lies less in any single technique than in the demonstration that security and efficiency need not be opposing forces. By combining established cryptographic primitives, secret sharing and partial homomorphic encryption, with machine learning optimizations like gradient sparsification and adaptive scheduling, and by engineering the whole into a deployable software system, Bao and Su have sketched what practical privacy-preserving AI might look like. As cognitive computing continues to penetrate healthcare, finance, and government, frameworks of this kind may become the invisible infrastructure that lets intelligent systems learn from humanity’s most sensitive data while keeping that data exactly where it belongs: out of sight.

Subject of Research: Privacy-preserving collaborative machine learning using encryption-based training and data protection systems

Article Title: Collaborative encryption-based machine learning model training method and data protection software system construction

Article References: Bao, D., & Su, W. (2026). Collaborative encryption-based machine learning model training method and data protection software system construction. Neural Computing and Applications, 38(19), Article 770. https://doi.org/10.1007/s00521-026-12383-7

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12383-7

Keywords: privacy-preserving machine learning, homomorphic encryption, Shamir secret sharing, federated learning, gradient sparsification, data privacy, secure multi-party computation, Transformer models, ResNet, adaptive scheduling, encrypted training, data protection