The review’s central claim is that the value of an algorithm cannot be understood from its mathematical form alone. It must be interpreted in relation to data type, deployment environment, computational budget, interpretability requirements, and operational risk. To establish this, the authors adopted a PRISMA-guided screening workflow, searching major academic databases including Springer, IEEE Xplore, ScienceDirect, Google Scholar, MDPI, Frontiers, Elsevier, Wiley, and arXiv. Their Boolean search combined AI method terms with domain and application terms and functional task terms, ensuring that retrieved studies actually applied machine learning in one of the four target domains for a specific analytical purpose rather than merely mentioning it in passing. After title and abstract screening against predefined eligibility criteria, followed by full-text review, 220 studies met all criteria and were used as the bibliography, with 221 studies forming the corpus from which prominent algorithms were selected empirically rather than from a general list found in the AI literature.

The resulting taxonomy is organized by functional role rather than by domain. In IoT security, the most prevalent algorithmic functions are traffic classification, intrusion detection, anomaly discovery, botnet recognition, and lightweight edge inference. In network optimization, the functions shift to routing, load balancing, channel allocation, congestion mitigation, and policy adaptation. Image processing demands feature extraction, segmentation, compression, visualization, object recognition, and global-context modelling. Smart agriculture requires yield prediction, disease detection, crop recommendation, weather forecasting, soil monitoring, irrigation support, and farm-level decision support. This dual-axis structure, cross-referencing each algorithm by both model family and typical role across domains, is the review’s most distinctive contribution, and it reveals that algorithmic depth per domain is unequal: IoT security and image processing span both classical algorithms such as SVM, KNN, and Random Forest alongside deep architectures, while smart agriculture is more evenly spread across tree-based and regression methods.

In IoT security, the review finds that classical supervised algorithms remain important even though network-flow features are structured and labeled data are available. SVM, KNN, Naive Bayes, Random Forest, and XGBoost are routinely used to classify traffic as benign or malicious, but these models become problematic when new attack types arise or traffic distributions shift. Unsupervised and semi-supervised methods partially overcome these drawbacks. Autoencoders, which reconstruct normal traffic and flag high reconstruction errors as anomalies, have become the most popular unsupervised paradigm for IoT intrusion detection, particularly useful for zero-day and novel attack detection where labeled data are scarce. Isolation Forest efficiently detects outliers in high-dimensional data. The review also documents a clear architectural trend: CNN-based intrusion detection is moving from single networks to hybrids combining spatial feature extraction with sequential or attention-based techniques. Studies cited include a CNN-GRU model achieving high detection accuracy across IoT datasets and a CNN-BiLSTM-Transformer explicitly addressing long-sequence dependencies on CIC-IDS2017 and BoT-IoT benchmarks. Crucially, the review notes that higher model complexity does not provide proportional detection gains but does cause computational cost to rise disproportionately, a finding with direct implications for edge deployment on resource-constrained devices.

In computer networks and ad hoc routing, the review identifies reinforcement learning as the natural framework for sequential decision-making under dynamic conditions. Networks must deal with congestion, mobility, interference, link failure, changing traffic demand, and quality-of-service requirements, and traditional routing protocols based on fixed metrics or predetermined rules struggle to adapt. Q-Learning, SARSA, Deep Q-Networks, Actor-Critic methods, and Policy Gradient techniques learn routing or resource-allocation policies by interacting with a network environment. The review traces Q-Learning’s expansion into multi-objective distributed learning for IoT routing, Deep Q-Networks’ integration with the OLSR routing protocol for dynamic MANETs, and Soft Actor-Critic architectures within digital-twin-aided vehicular edge computing frameworks that jointly optimize task offloading, transmit power, and resource block allocation. Graph Neural Networks emerge as a crucial concept because communication networks are naturally graph-structured, with nodes, links, capacities, delays, and traffic demands being relational rather than tabular features. The review cautions that reinforcement-learning models can be difficult to stabilize before training converges, that exploration is risky in operational networks, and that deep models add latency and computational overhead, so AI routing must be compared against robust protocol baselines and tested on unseen topologies with safety mechanisms to prevent unstable or unfair routing decisions.

In image processing and computer vision, the review documents the shift from handcrafted features such as edges, texture descriptors, color histograms, and shape features to data-driven deep learning. Convolutional Neural Networks remain the go-to choice for computationally limited and real-time applications due to their spatial inductive bias and parameter efficiency, even as Vision Transformers become a new architectural contender. Vision Transformers treat image patches as sequences and use self-attention to capture long-range dependencies and global contextual relationships, yielding impressive results in medical image segmentation, remote sensing classification, and plant disease localization where global context is essential. The review highlights token sparsification approaches that remove redundant tokens while retaining discriminative information, lowering calculation and memory utilization for deployable transformer-based vision systems in resource-constrained settings like autonomous vehicles and drones. Autoencoders continue to serve as powerful representation learning tools for denoising, compression, and anomaly detection, while PCA and t-SNE remain valuable for dimensionality reduction, feature compression, and visualization of deep learning feature distributions. The review notes that PCA can compress transformer-based visual representations by over an order of magnitude with very little drop in classification accuracy.

In smart agriculture, the review emphasizes that agricultural systems are affected by weather, soil type, crop variety, disease pressure, pest activity, water availability, farm practice, and seasonal variability, making them fundamentally different from purely digital domains. Models must combine heterogeneous data from field sensors, weather stations, satellite images, UAV images, soil analysis, and farm bookkeeping. Classical machine learning remains relevant because many agricultural datasets are structured and of moderate size: Decision Trees formulate interpretable rules for crop decision-making, Random Forest and XGBoost capture non-linear relationships between weather, soil, vegetation indices, and management variables for yield prediction, and Logistic Regression provides transparent probabilistic baselines for disease-risk decision support. Deep learning is particularly beneficial where agricultural information is visual or sequential, with CNNs used extensively for plant disease detection, weed recognition, and remote-sensing interpretation, and LSTMs applied to weather forecasting, crop growth monitoring, and irrigation scheduling. The review stresses that good performance on images from one region does not assure good performance in another because of differences in light, cultivar, disease expression, background, or camera quality, necessitating field validation and explainability linked to agronomic variables.

Across all four domains, the review identifies recurring cross-cutting gaps that collectively reveal where current AI adoption falls short. Dataset realism is the first: lab traffic, known attack scripts, curated images, and location-specific agricultural datasets rarely capture deployment diversity, so benchmark performance must be clearly differentiated from deployment readiness. Model drift is the second: IoT devices evolve, attackers evolve, farms experience seasonal and climate changes, and network traffic shifts with applications, yet static train-test validation is insufficient to assess reliability over time, demanding online learning, periodic retraining, drift detection, and human-in-the-loop review. Explainability is the third: security analysts must understand what traffic caused an alert, network operators should know why a learned routing policy altered traffic paths, and agronomists require explanations connected to visible symptoms, weather variables, or soil conditions. Resource-aware deployment is the fourth: deep models and reinforcement-learning policies can be costly on IoT edges, farm sensors, and mobile devices, so pruning, quantization, knowledge distillation, and edge-cloud partitioning deserve greater attention. The review argues that a slightly less accurate but faster, more stable, explainable, and energy-efficient model may be better in real deployments.

The review concludes with a practical algorithm-selection framework that begins with the shape of the data rather than model popularity. When data is well-structured, tabular, and not too large, Decision Trees, Random Forests, Gradient Boosting, XGBoost, Logistic Regression, and SVM are strong first-line options. For images, CNNs and Vision Transformers are suitable; for sequences, LSTMs, Transformers, and temporal CNNs; for graphs, Graph Neural Networks; for action-oriented sequential problems, reinforcement learning after simpler baselines have been achieved. Label availability is the second decision point: when labeled examples are plentiful, supervised learning is typically appropriate, but when labels are scarce or costly, unsupervised and semi-supervised methods such as autoencoders, Isolation Forest, DBSCAN, and variational autoencoders become appealing. Deployment context is the third: when interpretability is required, linear models and tree ensembles with feature-attribution techniques may be more suitable; when data ownership prevents centralization, federated learning might be appropriate; when immediate response is needed, latency and failure behaviours should be checked before adding model complexity. Validation design is the fourth: security systems demand adversarial robustness and false-positive cost analysis, routing requires topology variation and traffic bursts, computer vision needs external databases and real-world image variation, and smart agriculture requires geographic, seasonal, and crop-variety diversity. Without such validation, the ranking of algorithms may be misleading. The review’s ultimate message is that AI adoption in connected environments is additive rather than substitutive: classical supervised learning, ensemble learning, unsupervised anomaly detection, deep neural networks, reinforcement learning, graph learning, and federated learning all play special roles, and future research should focus on lightweight, robust, multimodal foundation models and explainable AI validated on realistic datasets across IoT, campus networks, smart agriculture, image processing, and ad hoc routing environments.

Subject of Research: Artificial Intelligence

Article Title: A cross-domain review of artificial intelligence techniques for iot security, computer networks, image processing and smart agriculture in connected environments

Article References: Sadiq, B. O., Buhari, M. D., Ibrahim, A., Edozie, E., & Shabiluh, A. (2026). A cross-domain review of artificial intelligence techniques for iot security, computer networks, image processing and smart agriculture in connected environments. Discover Artificial Intelligence, 6(1), Article 1535. https://doi.org/10.1007/s44163-026-02500-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02500-9

Keywords: cross-domain AI review, IoT security, computer networks, image processing, smart agriculture, machine learning, reinforcement learning, federated learning, connected environments, Discover Artificial Intelligence, cross-domain, review

The team, led by Maryam Mehrnezhad of Royal Holloway together with Thyla Van Der Merwe and Michael Catt, set out to answer three questions: what gaps exist in applicable laws when it comes to female-related data; how do FemTech systems misuse those gaps, whether intentionally or not; and how do these systems violate the regulations that do apply. They chose the United Kingdom and Switzerland as study sites because neither is an EU member, yet both have substantial business operating within the European Economic Area, making them subject to EU rules such as the General Data Protection Regulation while also governed by their own domestic frameworks.

The regulatory review covered four main instruments: the GDPR, the Swiss Federal Act on Data Protection, the United Kingdom’s Medical Devices Regulations 2002 overseen by the Medicines and Healthcare products Regulatory Agency, and the EU’s Regulation 2017/745 for medical devices. The researchers manually searched each law’s public documents for FemTech-related keywords including period, fertility, pregnancy, abortion, menopause, ovulation, lactation and hormone. The result was striking: none of the regulations mentions FemTech data directly. The GDPR does protect special category data such as health, sex life and sexual orientation, and the Swiss law protects data touching the intimate sphere, but the researchers found no specific data protection rules for FemTech data collected and processed outside clinics and medical settings.

The medical device regulations fared no better. The UK’s Medical Devices Regulations 2002 is more than two decades old and contains nothing on the security and privacy aspects of modern connected products, let alone FemTech-related data. The EU’s 2017/745 regulation does classify devices for the control or support of conception as medical devices, but its risk management provisions address safety only, with no mention of security or privacy risks. The researchers note that the MHRA is developing a new Software and AI as a Medical Device Change Programme, including a work package on cyber-secure medical devices, and argue that such efforts are needed immediately.

To see how companies behave within this patchwork, the team purchased ten off-the-shelf products spanning the sector’s categories, from the Elvie smart breast pump and Daysy cycle computer to the Perifit kegel trainer, the Frida by Vibio sex toy, the Livia menstrual pain reliever, the Balance menopause app and the Bellabeat Ivy health tracker. Between September 2022 and April 2023, two authors independently set up each device and its Android app on Google Pixel 6 phones, logged what data was requested or automatically gathered, and repeated any inconsistent experiments jointly. They then audited the products’ websites and apps for privacy notices, consent mechanisms and trackers.

The data inventory was extensive. User data included names and contact details, lifestyle information such as weight, diet and sleep, cycle length and ovulation days, pregnancy test results and due dates, nursing times and volumes, details of reproductive organs such as cervical mucus and muscle strength, sexual activities including contraceptive use, medical information such as medication types and lab report scans, and physical and emotional symptoms. More surprisingly, many systems also collected data about other people: nursing and sleep data about babies, social media profiles, and details of partners including names, ages, photos and partnered sexual activity. Several apps also demanded access to dangerous phone permissions, including the camera, microphone, storage, contacts, location and WiFi and Bluetooth radios.

On consent, the picture was damning. Under the GDPR and the ePrivacy Directive, consent to tracking must involve an unambiguous positive action, be separated from other matters, and offer genuine choices to accept or reject without nudging. The researchers found that all but one of the apps and websites studied appeared to violate these requirements. Most presented users with a privacy wall, forcing agreement to terms and a privacy policy before the service could be used, or highlighted the accept option while burying the alternative. The sole exception was the Balance Menopause App, which presented valid consent in its app, though its website did not. Privacy policies offered little reassurance: only four even referenced FemTech-related data, and those that did often contradicted themselves, such as one fertility app claiming to require explicit consent for health data while treating the mere submission of that data as the consent itself.

Tracking was pervasive. Using the Exodus Privacy audit platform, which statically analyses Android app code for tracker signatures, the researchers found between one and nine trackers embedded in each app. Using the privacy-focused Brave browser, they found that most product websites activated trackers before any user engagement with the cookie notice. One website, hidratespark.com, spawned more than seventy trackers while left open without interaction, and even attempted to access motion sensors when visited from a mobile device. By contrast, the Daysy cycle computer and Lady Comp fertility tracker each contained just one tracker, Google Tag Manager, showing that minimal tracking is technically feasible in this sector.

The researchers argue that these practices translate into differential harms for real people. Intimate data can be weaponised by ex-partners and family members through tech abuse and cyberstalking, particularly where devices offer shared remote modes. Employers and insurers represent further threat actors: FemTech data has reportedly been shared with third parties including employers, and insurance firms with access to family health histories could discriminate on the basis of inferred pre-existing conditions. Governments and political and religious organisations have also been identified as interested in such data, a concern sharpened by the overturning of Roe v. Wade in the United States and reports of data marketplaces selling information about period app users. Data breaches in this domain are especially serious because the data mixes health, biometric, genetic, sexual and even religious and political information, and marginalised groups have more to lose from disclosure.

The study also exposes how companies game the boundaries between legal categories. Apps that handle medical records can register on the Google Play Store under Health and Fitness, Entertainment or Tools categories rather than Medical, avoiding scrutiny; of the ten apps studied, only one was listed as medical. Companies often deliberately frame products as wellness or wellbeing offerings to escape medical device regulation, meaning cycle tracking apps reach the market largely unvetted even as the ovulation and pregnancy tests they accompany are regulated. The authors conclude that neither general data protection law nor medical device regulation, alone or combined, adequately protects users, and they call for domain-specific, sectorial regulation that explicitly acknowledges FemTech risks, alongside certification systems, compliance testing and penalties, with particular attention to marginalised user groups. Some movement is under way, including the UK regulator’s recent auditing project on FemTech apps and the EU’s proposed European Health Data Space, but the researchers warn that without explicit recognition of these technologies’ risks, users will continue to trade their most personal data for services that could be built to collect far less.

Subject of Research: Security, privacy and regulatory gaps in female-oriented health technologies (FemTech)

Article Title: Mind the FemTech gap: regulation failings and exploitative systems

Article References: Mehrnezhad, M., Van Der Merwe, T., & Catt, M. (2024). Mind the FemTech gap: regulation failings and exploitative systems. Frontiers in the Internet of Things, 3, Article 1296599. https://doi.org/10.3389/friot.2024.1296599

Image Credits: AI Generated

DOI: 10.3389/friot.2024.1296599

Keywords: FemTech, privacy, data protection, GDPR, medical devices, health apps, IoT security, tracking, consent, women's health, regulation, digital health

The review, led by Gideon Ogunniye, Amaya Hana, and Jeremy Watson of University College London’s Department of Science, Technology, Engineering and Public Policy, presents a case application of a socio-technical framework for IoT research and development in the United Kingdom. Its central claim is deceptively simple: IoT systems are highly dependent on data collection and sharing, constant connectivity, and remote control, which means their value is achieved through connected data that makes objects readable, recognisable, locatable, addressable, and controllable over a network. That same dependence creates privacy, ethics, trust, reliability, acceptability, and security issues for users and for those affected by the ecosystem, and the authors contend that technical solutions alone cannot adequately address these public concerns. The technical, social, cultural, and behavioural aspects of how we design, develop, manage, and evolve the IoT, they write, will be critical to its success.

The intellectual machinery behind the argument stretches back decades. Socio-technical systems theory holds that although technical and social subsystems are closely interrelated, they are distinct from each other: technical subsystems aim to achieve specified performance parameters, while social subsystems consist of human beings with unpredictable behaviour. The PETRAS framework follows an influential schema by Leavitt from 1965, developed through his experience of organisational change and focused on the relationships between people, tasks, structures, and technologies, later extended by Davis and colleagues in 2014 into a model of six interrelated elements embedded within an external environment. In this view, a work system has goals and metrics, people with varying attitudes and skills, a range of technologies and tools, a physical infrastructure, cultural assumptions, and processes and working practices, all operating within a wider context of regulation, stakeholders, and an economic environment whose importance varies with each system.

The framework’s real-world testbed is the PETRAS National Centre of Excellence for IoT Systems Cybersecurity, the world’s largest socio-technical research centre focused on the future implementation of the IoT. Its origins trace to a 2014 UK government Blackett Review on the Internet of Things, which identified socio-technical challenges concerning privacy, trust, and security that needed to be addressed to release market opportunity for the technology. The PETRAS research hub, established in 2016 and jointly funded by the UK Engineering and Physical Sciences Research Council and user partners, was a consortium led by University College London with Imperial College London, the University of Oxford, Warwick University, and Lancaster University. It brought together 12 UK academic institutions, 51 projects, six medium-scale demonstrators, and more than 110 user partners from industry, government agencies, and NGOs. In 2019, the PETRAS Centre was created under the Strategic Priorities Fund, growing to 24 UK research institutions, over 120 user partners, and 63 projects.

On the technical side, the Centre concentrates on how the interactions between IoT, artificial intelligence, and machine learning produce cybersecurity challenges that must be addressed if society and the economy are to harness their full potential benefits. According to the IBM Institute for Business Value, the full potential of IoT can only be realised with the introduction of AI, and machine learning, deep learning, and artificial neural networks are used to analyse and learn from collected data to create public services and values. Several PETRAS projects have modelled AI-based anomaly detection for consumer and industrial IoT: the ELLIOT project developed an early anomaly detection model that can be applied to industrial processes to detect early signs of cyber-attacks, while the PSWaRMS project developed an AI-based randomised target defence approach to reduce the risk of severe consequences from attacks aimed at IoT assets. Other projects investigated how AI techniques enable ubiquitous data collection, from machine learning behaviour patterns for driver profiles in autonomous vehicles to the PPIEM project’s implementation of ML techniques to evaluate the accuracy of occupancy estimates and future carbon dioxide predictions in smart buildings.

But the convergence of these technologies also introduces new privacy, security, and ethical issues, since threat actors may exploit AI to identify vulnerabilities and attack surfaces in IoT systems. The Centre’s answer is to combine social and technical expertise in the co-creation of new solutions, approaches, tools, and techniques that address both kinds of challenges simultaneously. Its goals include providing strategic advice and policy insight as the go-to resource on securing digital technologies at the periphery for the public and private sector, increasing the early adoption of new methods and technologies by its industry, service, and government user partners, and functionally improving the capacity of government to rise to the challenge of an arms race of ever more complex socio-technical systems and threats. The Centre operates in responsive pull and proactive push modes, co-creating immediately useable research with user partners while maintaining the capacity for fundamental challenges and emerging issues not yet apparent to partners.

The framework’s social subsystem is equally deliberate. The Centre is designed to accommodate transdisciplinary research spanning physical and behavioural sciences, bringing together experts from business schools, policy and governance programmes, and computer science to research new market models for raw and inferenced IoT data that may be mediated through technologies such as distributed ledger technologies, standards, and social incentive schemes. Its people element comprises governing components, including a governing board with industrial advisory, public sector, user research, and research excellence sub-groups, and administrative components such as the management team and an ethics review sub-group, working in synergy with an operations group of principal investigators, co-investigators, and research partners. The management team provides a single point of contact and works with academic researchers to ensure that research outputs are synthesised and communicated in a way that is translatable to industry, government, and broader non-technical audiences. The Centre’s culture of agility and collaboration has generated over 467 peer-reviewed articles in its publication database.

Perhaps the most vivid illustration of the framework in action is the P-PITEE project, which tackled a problem most city dwellers never notice: when IoT sensors are deployed in public spaces, the technology might not be visible at first sight to local residents and community groups. In partnership with Lancaster City Council, the project used design methods to develop new policies for transparent and ethical deployment of secure IoT sensors in public spaces, covering the collection, use, and sharing of data, considerations of data transfer versus edge processing, and cybersecurity questions related to data storage and sharing. The team developed a novel walking workshop approach, Taking IoT for a walk, in which members of the public were guided on a walkshop around Lancaster, encountering real and fictional instances of IoT deployment and discussing the benefits, risks, and language. The method, funded through the UK Economic and Social Research Council’s Festival of Social Sciences and later EPSRC Telling Tales of Engagement funding, spread to Leeds, East Riding, Dorset, and the London Borough of Hounslow, and the project launched an interactive tool called TrustLens to help organisations use technology in public spaces.

The review also maps the socio-technical requirements that any IoT deployment must satisfy, organised around the themes that give PETRAS its acronym: Privacy, Ethics, Trust, Reliability, Acceptability, and Security. On adoption and acceptability, a nationally representative survey of over 2,000 UK smart home consumers, built on the Unified Theory of Acceptance and Use of Technology 2, revealed that trust is fundamental to consumer technology in which transmission of personal and sensitive information is involved, and preliminary findings suggested that the most significant way to make IoT more acceptable is by reducing anxieties stemming from user concerns about privacy, trust, and security. On safety and security, findings point to the need for Security Ergonomics by Design, implementing safety and security requirements from the design phase throughout the lifetime of a device, dynamic risk assessment to cope with the dynamic nature of IoT, and study of emerging security and trust economics. On privacy and trust, a meta-analysis of social science and humanities research identified wide-ranging ethical issues including control and oversight of data flows, the balance between authentication and privacy, trade-offs between identification and privacy, the limits of the GDPR, transparency and auditing of autonomous and machine learning algorithms, and consent mechanisms, with trust negotiation, identity management, and fine-grained but not burdensome access control identified as technical trustworthiness requirements.

The governance picture is sobering. A review conducted with BSI and the Department for Digital, Culture, Media and Sport identified a fragmented and complex regulatory landscape in which IoT security standards have been developed by industry consortia with varying and possibly conflicting motivations, difficulty in monitoring the adoption and effectiveness of standards, and difficulty in establishing a security baseline across all application domains. The IoT Multidisciplinary Standards Platform project, collaborating with BSI, the Institution of Engineering and Technology, Digital Catapult, and GCHQ, surveyed the main barriers to navigating, adopting, and implementing relevant IoT standards, informing further investigation into the key challenges that small and medium-sized enterprises face in designing secure IoT products. The Centre’s influence reaches into legislation: its report on the UK Code of Practice for Consumer IoT Security was highlighted in a parliamentary debate on an amendment to strengthen cybersecurity in children’s products in the Product Security and Telecommunications Infrastructure Bill. The authors argue that the strategy of co-designing IoT research with critical stakeholders is a balanced approach that can help set more realistic expectations for all stakeholders, with significant implications for eliciting socio-technical requirements and for addressing public concerns about acceptability and adoption, safety and security, privacy and trust, standards, governance, and policy, and harnessing the economic benefits of IoT. For anyone developing new research centres or seeking to reinvigorate existing ones, the PETRAS experiment offers a guide: build the synergies between collaborators in from the start, and treat the humans as part of the system, not an afterthought.

Subject of Research: Internet of Things

Article Title: PETRAS: a socio-technical framework for Internet of Things research and development

Article References: Ogunniye, G., Hana, A., & Watson, J. (2024). PETRAS: a socio-technical framework for Internet of Things research and development. Frontiers in the Internet of Things, 3, Article 1336564. https://doi.org/10.3389/friot.2024.1336564

Image Credits: AI Generated

DOI: 10.3389/friot.2024.1336564

Keywords: Internet of Things, IoT security, socio-technical systems, PETRAS, cybersecurity, privacy, trust, AI, socio-technical, framework, Internet, Things

The core insight behind the research is that neither blockchain nor machine learning alone can solve the security problem of wireless sensor networks. Machine learning models can flag suspicious behavior, but their alerts can be spoofed or tampered with if the underlying data store is compromised. Blockchain provides immutable, tamper-resistant logging, but it cannot by itself decide what constitutes a threat. Elsayed’s framework therefore couples an unsupervised anomaly detection ensemble with a lightweight blockchain infrastructure that uses SHA-256 cryptographic hashing for secure and immutable data storage, so that every confirmed threat is written into a ledger that attackers cannot quietly rewrite.

The machine learning side of the system is an ensemble of three unsupervised techniques: K-Means clustering, Isolation Forest, and Autoencoder-based reconstruction. K-Means groups sensor behavior into clusters so that outliers fall outside normal groupings. Isolation Forest isolates anomalous samples by randomly partitioning the feature space, a method that naturally excels at spotting rare events. The Autoencoder, a compact neural network trained to compress and reconstruct normal traffic patterns, produces a high reconstruction error when confronted with data it has never seen before, which serves as a powerful anomaly signal. Because all three methods operate without labeled attack data, the ensemble can detect novel attacks that signature-based systems would miss entirely.

One of the most inventive elements of the framework is a localized majority-voting consensus mechanism introduced to validate detected threats and reduce false-positive alerts without imposing significant communication overhead. When a node flags an anomaly, nearby nodes vote on whether the observation is consistent with their own readings. Only when a majority agrees is the threat confirmed and logged. This approach sidesteps the energy-hungry consensus protocols of conventional blockchains, which would quickly drain the batteries of sensor nodes, and instead borrows the logic of Byzantine fault tolerance in a form light enough for tiny devices.

The system operates in three main phases. First, behavioral data collection and preprocessing gathers traffic and node statistics and prepares them for analysis. Second, real-time unsupervised anomaly detection runs the ensemble on each new sample as it arrives. Third, decentralized threat confirmation with tamper-resistant logging writes validated alerts into the blockchain. This pipeline ensures that detection is not only fast but also verifiable, since the alert record is cryptographically chained to previous entries and cannot be altered retroactively.

Extensive simulations on three benchmark datasets, WSN-DS, NSL-KDD, and UNSW-NB15, demonstrate excellent detection performance. The framework achieved an accuracy of 99.83 percent with a low root mean square error of 0.3940, and it provides robust defense against DoS and DDoS flooding, data injection, and node compromise attacks. These are precisely the attack classes that plague wireless sensor networks, where an adversary can jam a channel, inject false readings, or physically capture a node and extract its cryptographic keys.

Perhaps the most consequential results concern overhead, the perennial obstacle to deploying sophisticated security on tiny devices. The framework introduces only 0.92 to 1.48 millijoules of additional energy per node per epoch, maintains a compact memory footprint of 2.1 to 3.8 kilobytes per node, and achieves ultra-low inference latency of 12.93 to 19.4 microseconds per sample, with end-to-end alert logging under 1.15 milliseconds. These figures highlight the system’s practicality for long-term deployment in resource-constrained wireless sensor networks, where every joule and every byte matters.

The significance of this work extends beyond the laboratory. As smart infrastructure, industrial IoT, and environmental monitoring expand into billions of connected sensors, the attack surface grows in tandem, and conventional perimeter defenses are useless for networks scattered across fields, pipelines, and cities. A decentralized, tamper-resistant detection layer that runs on the devices themselves offers a fundamentally different security model, one in which the network defends its own integrity without relying on a central authority that could become a single point of failure.

Challenges remain before such systems reach widespread adoption. Real-world deployments face radio interference, node churn, and heterogeneous hardware that simulations cannot fully capture, and the energy cost of blockchain logging, though minimal in this design, accumulates over years of operation. Yet the study’s demonstration that machine learning detection and blockchain verification can coexist within a few kilobytes of memory and a few millijoules of energy per epoch marks a meaningful step toward securing the invisible nervous system of the Internet of Things, one sensor at a time.

Subject of Research: Computer Science

Article Title: A Synergistic approach: blockchain-integrated machine learning based intrusion detection system for improving wireless network security

Article References: Elsayed, W. M. (2026). A Synergistic approach: blockchain-integrated machine learning based intrusion detection system for improving wireless network security. Neural Computing and Applications, 38(20), Article 812. https://doi.org/10.1007/s00521-026-12503-3

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12503-3

Keywords: wireless sensor networks, intrusion detection system, blockchain, machine learning, anomaly detection, IoT security, cybersecurity, Synergistic, approach, blockchain-integrated, machine, learning

The research, published in Frontiers in Big Data, presents IOTTRUST, a graph-augmented intrusion detection framework developed by Nachaat Mohamed of Rabdan Academy and NEUSAI Technologies in the United Arab Emirates and Hamed Taherdoost of University Canada West and several affiliated institutions. Rather than replacing conventional classifiers with a full graph neural network, IOTTRUST constructs a directed, weighted network flow graph from the widely used UNSW-NB15 benchmark and enriches per-flow machine learning with twenty-four graph-derived topology features computed for each source and destination IP. The approach delivers competitive accuracy with transparent, operationally meaningful features and without GPU-based neural network training, making it substantially more accessible for deployment on resource-constrained IoT gateways.

The graph itself is strikingly small and interpretable. From the full UNSW-NB15 corpus of 440,047 flow records, the construction yields just 47 IP nodes and 438 directed edges, organized into ten /24 subnet communities. Community structure analysis reveals a dramatically skewed world: four attacker IPs in the 175.45.176.x subnet account for 89.85 percent of all attack flows, while the remaining nine communities contain exclusively normal traffic. The top attacker address, 175.45.176.0, shows an individual attack ratio of 0.9632, meaning 96.3 percent of its 38,614 outgoing flows are attack traffic, and it targets eleven distinct destination IPs spanning both victim and server subnets. Attacker out-degrees run two to three orders of magnitude larger than those of normal hosts, and attacker fan-out ratios of 0.23 to 0.26 dwarf the zero fan-out of server IPs that only receive traffic.

What distinguishes this study from much of the graph-based security literature is its unusually honest evaluation protocol. An earlier version of the work computed graph statistics from the full dataset, including flows the classifier was later evaluated on, introducing temporal leakage; it also compared models trained on different dataset sizes. The revised protocol addresses both weaknesses simultaneously. A stratified sample of 40,000 flows is drawn with a fixed random seed, split 70/30 into training and test partitions before any graph statistic is computed, and all per-IP topology features are derived exclusively from the 28,000-flow training partition, then propagated to test flows without any access to test-set labels. All three experimental tracks are evaluated on this identical, size-matched, leakage-free sample.

Under this stricter protocol, the headline findings reverse in an informative way. A Flow-Only model using the standard 43 UNSW-NB15 features achieves 98.98 percent accuracy with a false positive rate of 0.81 percent. A Graph-Only model trained exclusively on the 24 topology features reaches 97.89 percent accuracy with a false positive rate of 1.79 percent, meaning graph features alone are not sufficient to match the full flow representation. But the full IOTTRUST Hybrid model, combining 22 numeric flow features with the 24 graph features, achieves the best result on every metric: 99.07 percent accuracy, an AUC of 0.9996, precision of 97.09 percent, recall of 98.30 percent, and a false positive rate of 0.74 percent. The improvement over Flow-Only is modest in absolute terms but statistically significant across five-fold cross-validated F1-scores (p = 0.026), while the advantage over Graph-Only alone is large and highly significant (McNemar p < 0.001).

Feature importance analysis confirms that network relational structure continues to carry predictive information even under the leakage-free evaluation. Twelve of the top fifteen most important features in the Hybrid model remain graph topology features, accounting for 60.4 percent of the top-15 importance mass. The single most important feature is the historical attack ratio of the destination IP, computed from the training graph, followed by a raw flow feature, the source-to-destination time-to-live, and then the source-side attack ratio. The mechanism is interpretable: the four attacker IPs maintain high historical attack ratios of 0.64 to 0.96 that are structurally stable across the training partition, so any flow whose endpoints carry a non-zero attack ratio is overwhelmingly likely to belong to the attack community, providing the Random Forest with a near-deterministic signal before it examines any per-flow byte count or timing statistic.

The residual errors are not randomly distributed but concentrated where the topological signal is weakest. All 42 false negatives originate from the attacker subnet, and 32 of them are Fuzzers, a category whose per-flow byte and timing statistics overlap substantially with normal traffic even when the source IP’s history is known. Conversely, all 68 false positives are flows from attacker IPs that the benchmark labels Normal, a known artifact of the IXIA PerfectStorm traffic generator, which produces both attack and benign traffic from the same physical machines. The graph-topology model correctly flags these flows as suspicious based on the source address’s behavioral history, suggesting that the true false positive rate for genuinely benign flows from genuinely benign IPs may be lower than the label-based figure implies.

The operational significance lies in alert fatigue. In a production IoT deployment processing 500,000 normal flows per hour, a system at 0.81 percent false positive rate would generate roughly 4,050 false alerts per hour; IOTTRUST at 0.74 percent would generate roughly 3,700, a reduction of about 350 analyst alerts per hour, achieved alongside a simultaneous improvement in recall. The computational overhead is negligible: constructing the training graph from 28,000 flows takes 27.3 milliseconds on a single CPU thread, construction from the full 440,047-flow dataset takes about half a second, and per-flow feature lookup is a constant-time dictionary access of roughly 4.77 microseconds, independent of graph size. Ablation studies show that degree and structural features account for the majority of the gain, while the attack-ratio features, though top-ranked individually, are strongly correlated with degree statistics on this dataset.

The framework also offers a label-free deployment path. Because the attack-ratio features depend on historical labels, the authors evaluated an Isolation Forest-based unsupervised proxy, aggregating per-flow anomaly scores by IP address. The resulting model achieves 99.08 percent accuracy and a false positive rate of 0.73 percent, essentially indistinguishable from the supervised variant, indicating that IOTTRUST can operate in environments where no labeled historical traffic is available. In operational settings, the supervised features correspond to maintaining a periodically updated IP reputation table, recomputed from flows that security analysts have confirmed as benign or malicious, directly analogous to established threat-intelligence practice.

The authors are candid about limitations. The 47-node graph is far smaller than production IoT networks, though the linear-time construction and constant-time lookup scale tractably; a million-node statistics dictionary would consume roughly 96 megabytes of memory, well within the capacity of a standard IoT gateway. Subnet-based community assignment works well on UNSW-NB15, where administrative structure mirrors attack topology, but in NAT-heavy or heterogeneous enterprise networks the Louvain algorithm, which derives communities purely from observed connectivity, is preferable; on this dataset Louvain-based assignment yields a marginal but consistent improvement, reaching 99.12 percent accuracy and a 0.68 percent false positive rate. Validation on newer datasets such as TON-IoT and CIC-IoT2023, where attacks are distributed across many device IPs rather than concentrated in four dedicated attacker nodes, remains the highest-priority future work, and adversaries aware of graph-based detection could attempt to disguise their topology signatures by rotating IPs or diluting their attack ratios. Within these caveats, the study makes a defensible, reproducible case that network relational structure provides a genuine, if incremental, enrichment to per-flow intrusion detection, and that who is involved in a flow, and how that party behaves across the entire network, is a fundamentally different and orthogonal signal to what the flow contains.

Subject of Research: Cybersecurity

Article Title: IOTTRUST: graph-based anomaly detection for IoT intrusion using network flow topology and community structure analysis on UNSW-NB15

Article References: Mohamed, N., & Taherdoost, H. (2026). IOTTRUST: graph-based anomaly detection for IoT intrusion using network flow topology and community structure analysis on UNSW-NB15. Frontiers in Big Data, 9, Article 1885965. https://doi.org/10.3389/fdata.2026.1885965

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1885965

Keywords: IoT security, intrusion detection, graph-based anomaly detection, UNSW-NB15, network topology, community detection, machine learning, false positive reduction, IOTTRUST, graph-based, anomaly, detection

Modern intrusion detection systems face a paradox that has shaped the field for years. Signature-based detectors are precise but blind to zero-day attacks, while representation-learning approaches built on deep neural networks generalize better yet often behave as opaque black boxes whose decisions cannot be audited. Federated methods promise collaborative learning across sites without centralizing sensitive traffic data, but in most published systems these three paradigms operate independently, each compensating poorly for the weaknesses of the others. The result, the authors argue, is a landscape of detectors that lack causal interpretability, waste feature information, crumble under adversarial pressure, and fail to generalize when deployed across multiple sites.

The new framework, developed by Subham Kumar Gupta and Yogesh Kumar Gupta of Banasthali Vidyapith and Atul Kumar Verma of Manipal University Jaipur, is built from five cooperating components. The Causal-Adaptive Signature Learning Engine, or CASLE, constructs a causal signature graph derived from structural causal models, allowing the system to distinguish genuine cause-and-effect relationships in network traffic from mere statistical correlations that attackers can easily spoof. This causal grounding is what gives the detector its interpretability: instead of flagging traffic because it resembles a memorized pattern, the system can trace which upstream variables actually drive a downstream attack signature.

Sitting upstream of the causal engine is the Entropy-Guided Feature Compression Network, abbreviated EGFC-Net. Rather than feeding every available traffic attribute into the classifier, this module applies entropy optimization to identify and retain only the features that carry maximal information about attack behavior. In practice, the framework compressed the original feature space of each dataset down to just 18 retained features, a reduction of 70.9 percent. That pruning matters for more than elegance: fewer features mean lower computational load, faster inference, and less surface area for adversarial manipulation, all critical properties for detectors that must run at line speed on edge devices and Internet of Things gateways.

The compressed, causally filtered features then flow into the attention-based Hybrid Cognitive Detection Synthesizer, which produces a cognitive detection matrix fusing the complementary strengths of rule-based logic and learned representations. Attention mechanisms weight the contribution of each detection pathway dynamically, so that a burst of anomalous DNS traffic, for example, can tilt the fusion toward the features most diagnostic of command-and-control behavior. The synthesizer outputs robustness-calibrated detection scores, which are then adjusted by the Adversarial Robustness Calibration Module. This module was stress-tested against bounded gradient-based perturbations, including fast gradient sign method style attacks, and the framework held its detection accuracy above 91.5 percent, a notable result given how easily deep-learning detectors can be fooled by small, carefully crafted input perturbations.

The fifth component, the trust-weighted Federated Threat Intelligence Evolution Framework, addresses the deployment problem. In a federated setup, participating sites train locally on their own traffic and share only model updates, preserving data privacy while allowing a global adaptive threat intelligence model to evolve from the collective experience of the network. Trust weighting ensures that updates from sites with consistently reliable data contribute more strongly to the global model, a safeguard against both noisy participants and adversaries attempting to poison the shared intelligence. This architecture is aimed squarely at the environments where intrusion detection is hardest: cyber-physical systems, Internet of Medical Things deployments, unmanned aerial vehicles, and distributed IoT networks where no single site sees the full picture.

The evaluation methodology was deliberately conservative. The authors tested on three publicly available and structurally distinct datasets: CICIDS2017, a widely used benchmark of realistic enterprise traffic attacks; UNSW-NB15, which combines synthetic attack generation with real background traffic; and IoT-23, a corpus of IoT malware captures. Dataset-specific partitions were used to control for training leakage, five-fold cross-validation probed the model’s ability to learn generalizable patterns, and ten evaluations with different random seeds established stability. Across those ten iterations, the framework achieved an aggregate accuracy of 96.37 percent, precision of 95.97 percent, recall of 95.43 percent, and an F1-score of 95.68 percent.

Benchmarks against competing approaches were run under matched preprocessing conditions and identical class distributions, a detail that matters because unfair comparisons are endemic in the intrusion detection literature. The hybrid framework outperformed blockchain-based collaborative IDS designs, traditional machine-learning pipelines, and optimized XGBoost-based detectors. Inference latency averaged 18.6 milliseconds per decision, fast enough for real-time deployment on the network edges where IoT and medical-device traffic actually arrives.

What makes the approach conceptually interesting beyond its benchmark numbers is the way it distributes epistemic labor across specialized modules. Causal models supply explanations and resistance to spurious correlations; entropy compression supplies efficiency and a smaller attack surface; attention-based fusion supplies adaptability; adversarial calibration supplies robustness; and federation supplies collective learning without data centralization. Each component addresses a failure mode that has historically sunk single-paradigm detectors, and the ablation-style logic of the design suggests the authors built the architecture by asking, for each known weakness, which mechanism could compensate for it.

The work arrives as networks grow more distributed and adversarial machine learning becomes a practical threat rather than a laboratory curiosity. As connected devices proliferate through hospitals, factories, and critical infrastructure, the demand for detectors that are simultaneously accurate, interpretable, robust, and privacy-preserving will only intensify. A framework that treats those four properties as co-equal design constraints, rather than trading one against another, offers a template for what next-generation network defense might look like, even as real-world deployments will need to test whether laboratory benchmarks translate to the messy traffic of production networks.

Subject of Research: A causality-driven hybrid intrusion detection system using entropy-aware feature learning and federated threat intelligence for network attack detection

Article Title: Causality-driven hybrid intrusion detection with entropy-aware feature learning and federated intelligence

Article References: Gupta, S. K., Gupta, Y. K., & Verma, A. K. (2026). Causality-driven hybrid intrusion detection with entropy-aware feature learning and federated intelligence. Scientific Reports. https://doi.org/10.1038/s41598-026-69897-2

Image Credits: AI Generated

DOI: 10.1038/s41598-026-69897-2

Keywords: intrusion detection, causal learning, entropy-based feature selection, federated learning, adversarial robustness, cybersecurity, machine learning, IoT security, threat intelligence, deep learning, network security, Causality-driven

The core problem, according to the review by Kai Liang Lew, Kok Swee Sim, and Hao Xian Cheong, is what researchers call closed-set evaluation. Most intrusion detection models are trained and tested on the same dataset, under the assumption that every network flow they encounter will belong to one of the attack categories seen during training. That assumption collapses the moment a model meets live traffic. Real networks experience distribution shifts as devices come and go, protocols evolve, and traffic patterns change with the seasons. Worse, attackers constantly invent new techniques, meaning a deployed system will inevitably face attack families that were never in its training data. A model that cannot tolerate these conditions may look invincible in the lab and fail silently in production.

The first deployment gap the reviewers examined is cross-dataset generalization, the ability of a model trained on one collection of traffic to perform on another. The evidence here is sobering. One study reported that macro-F1 scores, which balance precision and recall across attack classes, fell from a range of 0.84 to 0.96 under in-domain testing to just 0.69 to 0.78 when the model was tested on a harmonized future internet IoT dataset, accompanied by a benign false-positive rate of roughly 0.30. That last figure deserves particular attention: because benign flows vastly outnumber malicious ones in real traffic, even a modest false-positive rate on normal traffic can flood security analysts with more false alerts than genuine detections. Another study found that while 9 of 16 within-dataset experiments achieved Matthews correlation coefficients above 99.6 percent, the average MCC plummeted from 94.63 percent to 29.35 percent under cross-dataset testing, a near-total loss of discriminative power.

Why do models that excel on their home turf fall apart elsewhere? The review identifies three culprits. First, the tools used to extract network features such as timeouts and packet counts apply different rules in different environments, so the feature data from two networks may look radically different even when the underlying behavior is similar. Second, deep networks are prone to shortcut learning: instead of learning generalizable signatures of malicious behavior, they latch onto benign traffic footprints specific to a dataset, including IP subnets, operating system artifacts, and active ports. Third, the alignment methods designed to bridge different data sources can trigger negative transfer, actively degrading accuracy when the source and target domains are too dissimilar. Researchers have tried to close these gaps with techniques like Wasserstein distance-guided feature alignment, few-shot multidomain fusion, and class-incremental adaptation, and these approaches do shrink the performance drop, but residual sensitivity to domain shift persists across the literature.

The second gap is open-set recognition, and it may be the most dangerous. A closed-set classifier is forced to assign every flow to one of its known labels, which means a genuinely novel attack, say a zero-day exploit, gets confidently mislabeled as something familiar or dismissed as benign. Some studies have begun addressing this. Souza and colleagues built an energy-based flow classifier with a dedicated suspicious class rather than forcing every flow into a known attack label, achieving an area under the receiver operating characteristic curve of 0.859 on a reproduced CICIDS2017 open-set benchmark. Others have combined unknown-attack handling with cross-dataset evaluation, testing whether rejection mechanisms survive when both the labels and the traffic distribution change simultaneously. Yet the review found that most open-set work still relies on single-dataset holdout protocols, leaving open the question of whether an open-set detector remains trustworthy when the network itself shifts underneath it.

The third gap is reliability, and it is perhaps the most counterintuitive. High accuracy does not guarantee that a model’s confidence is meaningful. A calibrated model’s stated probability should match its actual likelihood of being correct; an overconfident model that says 99 percent when it is right only 70 percent of the time is a liability, especially when automated responses hang on those numbers. The reviewed studies deployed an arsenal of reliability techniques: Bayesian deep learning replaces fixed weights with distributions so predictions carry uncertainty, Monte Carlo dropout runs dropout at inference time to estimate variability, and abstention methods allow a model to refuse a prediction when confidence is too low. One study using Bayesian autoencoder uncertainty achieved an area under the rejection curve of 0.986 on UNSW-NB15, while another’s Wisdomnet approach drove accepted-sample error to near zero at rejection rates of roughly 5 to 7 percent. But here again, the reviewers found a pattern: calibration and abstention were typically measured only on in-distribution traffic, and calibration can degrade badly when data shifts or new attacks appear, leaving the model overconfident precisely when it matters most.

The comparative analysis reveals a fragmented landscape. Of the 29 empirical works, some tackled cross-dataset transfer, some open-set rejection, some calibration and uncertainty, and some practical deployment concerns like federated learning, edge computing, and poisoning robustness. A handful combined two of the three core concerns: one Transformer-based intrusion detection system paired cross-dataset evaluation with calibration analysis, reporting AUC improvements of roughly 10 to 20 percent and F1 gains of more than 10 points over baselines, while open-set domain adaptation studies linked unknown-attack handling with cross-domain variation. But no study delivered all three under a single unified protocol. The reviewers also graded study quality, flagging issues like test-informed threshold selection, unclear data splits, missing repeated-run uncertainty, and inconsistent reporting, and their confidence ratings ranged from higher to limited depending on protocol clarity and leakage control.

Reproducibility adds another layer of concern. Training-testing splits, label harmonization rules, and held-out attack designs vary inconsistently across the corpus, making direct comparison between studies nearly impossible. Dataset coverage is narrow as well: most results rest on a small set of well-worn benchmarks such as CIC-IDS2017, UNSW-NB15, and NSL-KDD, while newer internet traffic remains underused. The reviewers caution that their reported numbers should be read as within-study evidence rather than rankings, since differences in feature harmonization, source-target pairing, openness regimes, and threshold selection confound any cross-study comparison. In other words, the field’s benchmark culture, optimized for chasing leaderboard accuracy, is structurally ill-suited to answering the question that actually matters: will this model protect a real network?

To fix this, the authors propose a joint deployment-oriented evaluation protocol built on a four-step pipeline. First, a model undergoes offline training and post hoc calibration to establish baseline confidence curves. Second, it runs inline at network gateways, predicting labels and emitting calibrated confidence scores for incoming flows. Third, dual-path decision routing sends high-confidence predictions to automation while routing low-confidence cases and open-set detections to human analysts or deep-packet inspection tools. Fourth, flagged anomalies are periodically labeled and fed back through class-incremental updates and recalibration cycles. The protocol also specifies fixed source-target dataset pairs, the withholding of at least one attack family from training with explicit reporting of openness ratios, and calibration and abstention metrics such as expected calibration error, Brier score, and risk-coverage curves computed on shifted and unknown-attack cases, not just familiar traffic. All thresholds would be selected using only training or validation data before held-out testing.

The message for the cybersecurity industry is blunt but constructive. The limitation of current AI intrusion detection is not weak models but weak evaluation: without cross-dataset testing, dataset-specific overfitting cannot be ruled out; without open-set testing, novel zero-day attacks get forced into known labels, creating security blind spots; and without calibration, models make overconfident, incorrect calls on shifted traffic. The three checks together form the minimum evidence base for deployment readiness, though even they do not capture everything, since operational readiness also involves latency, throughput, resource consumption, temporal drift, adversarial robustness, and analyst workload. Until the field adopts joint protocols that stress-test generalization, novelty, and reliability simultaneously, the eye-popping accuracy figures filling the literature will remain what the Malaysian team suggests they are: impressive numbers about a world that no deployed system will ever actually inhabit.

Subject of Research: Deployment-oriented evaluation of deep learning network intrusion detection systems

Article Title: Deployment-oriented evaluation of deep learning network intrusion detection systems

Article References: Lew, K. L., Sim, K. S., & Cheong, H. X. (2026). Deployment-oriented evaluation of deep learning network intrusion detection systems. Frontiers in Communications and Networks, 7, Article 1863367. https://doi.org/10.3389/frcmn.2026.1863367

Image Credits: AI Generated

DOI: 10.3389/frcmn.2026.1863367

Keywords: network intrusion detection, deep learning, cross-dataset generalization, open-set recognition, calibration, uncertainty quantification, IoT security, IIoT, software-defined networking, domain shift, zero-day attacks, cybersecurity evaluation

The new algorithm, described by Qian Wang, Jiahui Shang, Weihao Huang, Yongqiang Cheng and Bing Zhang, is a semi-supervised, domain-adaptive intrusion detection method built around a generative adversarial network that maximizes mutual information between features, combined with a careful scheme for choosing pseudo-labels. In plain terms, the system learns to detect attacks on a new network even when almost none of that network’s traffic has been labeled, by transferring what it knows from a labeled source domain while reshaping its internal representation so that the quirks of one dataset do not contaminate its judgments about another.

The core difficulty the authors address is known as domain shift. When an intrusion detection model is trained on, say, a public benchmark dataset and then confronted with traffic from an Internet of Things deployment or a private enterprise network, the distributions of the two datasets differ in countless small ways: different protocols, different device mixes, different attack patterns. Features that carried strong predictive signal in the source domain may be nearly meaningless, or even misleading, in the target domain. Earlier work on domain-adaptive intrusion detection has tried to align the feature distributions of source and target data, but aggressive alignment can erase information that actually matters for distinguishing attacks from benign traffic.

The team’s answer is to align distributions while preserving semantics, and their tool for this is a generative adversarial network augmented with a mutual information objective. Mutual information is an information-theoretic quantity that measures how much knowing one variable tells you about another. By maximizing the mutual information between the original features of the source and target domains and the domain-invariant features that the network extracts, the model ensures that the compressed, environment-agnostic representation it builds retains as much of the integrated semantic content of the raw data as possible. The result, the authors report, is a set of domain-invariant features that are effectively equivalent to the original features in information content, yet no longer betray which dataset they came from — precisely the property an intrusion detector needs when it must operate across environments.

Before that adversarial game even begins, the researchers tackle another chronic headache in intrusion detection: class imbalance. Real network traffic is overwhelmingly benign, and public attack datasets are often dominated by a handful of frequent attack types, so models trained naively overfit to majority classes and neglect rare but dangerous ones. The team proposes an improved oversampling method, LR-SMOTE, which builds on the Limit Radius Synthetic Minority Oversampling Technique previously developed using k-means clustering and support vector machines. Their enhancement generates synthetic minority-class samples using a multi-sample strategy with random smoothing, expanding the representation of underrepresented attack categories in the source domain before the distribution alignment step. This pre-balancing prevents the adversarial alignment from washing minority attack classes away in pursuit of overall distributional similarity.

Once the domain-invariant features are available, the algorithm exploits the sea of unlabeled target-domain traffic through pseudo-labeling — the practice of letting a trained model assign provisional labels to unlabeled data and then training on those labels as if they were ground truth. Pseudo-labeling is powerful but risky: a confidently wrong pseudo-label can propagate error through the model. To keep the feedback loop honest, the researchers adopt an Uncertainty-aware Pseudo-label Selection framework, which admits only those provisional labels the model assigns with sufficient confidence, filtering out ambiguous cases that would inject noise. The selected pseudo-labeled samples then train a One-class Support Vector Machine classifier over the domain-invariant features in a semi-supervised manner, allowing the detector to make full use of unlabeled target traffic rather than discarding it.

The architecture is notable for how each component compensates for a specific failure mode of the previous state of the art. Standard adversarial domain adaptation, popularized by the domain-adversarial training of neural networks approach introduced by Ganin and colleagues, aligns distributions but can sacrifice class-discriminative structure. Adding the mutual information maximization term, an idea drawn from self-supervised representation learning, anchors the alignment to the original feature content. Pseudo-labeling with uncertainty selection, in turn, closes the loop by letting the adapted representation be refined with target-domain supervision that costs nothing to produce. The combination means the system needs labeled data only in the source domain, which is exactly the asymmetry that real deployments exhibit.

The experimental evaluation spanned both conventional network intrusion datasets and Internet of Things datasets, testing the algorithm in cross-domain scenarios where training and testing data come from different distributions. The datasets referenced in the study include widely used benchmarks such as UNSW-NB15, CICIDS-style collections, and IoT-focused resources like Bot-IoT and TON-IoT, which capture the distinctive traffic profiles of connected devices. Across these experiments, the authors report that the proposed approach demonstrates effectiveness in cross-domain intrusion detection scenarios, validating that the mutual-information-constrained adaptation plus uncertainty-filtered pseudo-labeling outperforms the naive transfer of models trained purely on public data. The study builds on the team’s earlier work on domain-adaptive IoT intrusion detection using joint domain adversaries, extending it with the semi-supervised machinery that lets unlabeled target data do real work.

The practical significance is considerable. Security operations centers are chronically short of labeled incident data for their own environments, and the most sensitive networks — hospital systems, industrial control systems, smart buildings — are precisely those where curated public datasets are least representative. An algorithm that can harvest knowledge from labeled public benchmarks, adapt it to the idiosyncrasies of local traffic, and improve itself on the fly using unlabeled flows could dramatically lower the cost of deploying competent intrusion detection in resource-constrained settings, including IoT fleets where devices lack the compute to run heavyweight models and the cloud-side detectors they depend on must generalize across heterogeneous device populations.

The work also joins a broader research movement to make machine learning robust to the gap between laboratory benchmarks and the messy real world. Recent studies have compared synthetic and real-world intrusion detection datasets and found the discrepancy sobering, and researchers in adjacent fields — fault diagnosis of machinery, malware detection, smart-grid security — have pursued parallel domain adaptation strategies. By showing that mutual information maximization can keep adapted features semantically faithful while adversarial training strips away domain-specific artifacts, and that uncertainty-aware pseudo-label selection can safely convert unlabeled traffic into training signal, the Yanshan University team and their British collaborator offer a template that security researchers far beyond intrusion detection are likely to study. The work was supported by the Science and Technology Program of Hebei Province, and the authors state that the datasets analyzed are available from the corresponding author on reasonable request. As cyberattacks grow more automated, the defenses that guard against them may increasingly be algorithms that teach themselves — quietly, on unlabeled traffic — what an attack looks like in a network they have never seen labeled.

Subject of Research: Semi-supervised domain-adaptive network intrusion detection using adversarial mutual information maximization and pseudo-labeling

Article Title: A semi-supervised intrusion detection algorithm using mutual-information maximization adversarial network and pseudo-labeling

Article References: Wang, Q., Shang, J., Huang, W., Cheng, Y., & Zhang, B. (2026). A semi-supervised intrusion detection algorithm using mutual-information maximization adversarial network and pseudo-labeling. Cluster Computing, 29(15), Article 845. https://doi.org/10.1007/s10586-026-06643-3

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06643-3

Keywords: intrusion detection, domain adaptation, semi-supervised learning, generative adversarial network, mutual information, pseudo-labeling, IoT security, class imbalance, SMOTE, machine learning, cybersecurity, domain-invariant features

The team searched Google Scholar and IEEE Xplore for peer-reviewed studies published between 2021 and 2025, supplementing the search with manually incorporated papers indexed in Scopus and Web of Science and updated through early 2026. From an initial pool of 92 records, the researchers removed 14 duplicates, screened 78 titles and abstracts, excluded 30, retrieved 48 full texts, and discarded 16 more, ultimately including 32 studies in the final synthesis. Two reviewers performed screening and data extraction independently, resolving disagreements by consensus and consulting a third reviewer when needed. Because the included studies varied wildly in design, spanning experimental prototypes, simulations, conceptual frameworks, and prior reviews, the authors judged that a conventional statistical meta-analysis was impractical. Instead they employed a narrative, thematic synthesis, assessing each study’s methodological transparency, the adequacy of its evaluation metrics, the validity of its results, and its honesty about limitations.

The single largest body of work concerns smart home energy management systems, known as HEMS, which combine IoT sensors, consumption modeling, and optimization algorithms to schedule appliances, shave peak demand, and interact intelligently with the electricity grid. The reported numbers are striking. Studies in the review documented energy savings ranging from 15 to 25 percent under machine learning and rule-based control, while fuzzy logic controllers achieved savings of up to 70 percent under dynamic and uncertain load conditions. Artificial neural networks improved predictions of household consumption enough to refine lighting and HVAC control, and deep learning models, particularly long short-term memory networks, boosted load forecasting accuracy and demand response performance. Cloud-connected dashboards fed by smart meters, current transformers, and voltage sensors gave households real-time visibility into their consumption, sometimes through nothing more exotic than a messaging app, enabling remote appliance control and behavioral feedback.

Beneath these headline savings lies a more granular technique that the review treats as a field in its own right: non-intrusive load monitoring, or NILM. Rather than instrumenting every appliance with its own smart plug, NILM infers what individual devices are doing by analyzing the aggregate electrical signature measured at a single point, typically the main meter. Machine learning models extract power signatures and consumption patterns from waveform data, disaggregating the household total into appliance-level activity. This enables targeted interventions, such as identifying an inefficient refrigerator or revealing wasteful usage habits, without additional hardware. One AIoT-based system in the reviewed literature reported 35 percent energy savings alongside anomaly detection. Yet the authors found dedicated NILM research scarce, with the technique usually embedded invisibly inside broader monitoring platforms, and they flagged a particularly acute obstacle: the absence of large annotated datasets from African households, which limits the development and validation of models for the settings where such technology could matter most.

Reinforcement learning represents the most ambitious corner of the field. Here, a control agent learns by trial and feedback, continuously adjusting its decisions to balance comfort, cost, and efficiency under uncertainty. The reviewed studies demonstrated adaptive appliance scheduling, HVAC regulation that responds to shifting environmental variables, and dynamic load shifting that avoids expensive peak periods. One deep reinforcement learning system cut electricity costs by 31.5 percent while improving power factor, and a hybrid deep reinforcement learning framework paired with explainable AI improved water-energy efficiency in urban management. A separate study used proximal policy optimization, a state-of-the-art RL algorithm, to achieve adaptive appliance control. The catch, the authors emphasize, is that most of these successes live in simulation. Running learning agents on the resource-limited microcontrollers that populate real smart home prototypes remains difficult, and practical concerns such as safety during exploration, sample efficiency, and respect for ingrained user habits are rarely addressed in the literature.

Water, the review’s quieter protagonist, receives far less attention than electricity despite its growing scarcity. The smart water studies that made the cut relied heavily on IoT instrumentation: soil-moisture probes for irrigation, ultrasonic tank-level detectors, flow sensors, and leak detectors feeding real-time data to cloud dashboards and mobile interfaces. Some systems autonomously activated pumps when storage levels dropped, reducing both manual operation and pumping costs. More sophisticated approaches brought digital twins, virtual replicas of distribution networks, together with machine learning and reinforcement learning to optimize water treatment and predict consumption, while gamified feedback systems encouraged conservation behavior, particularly among younger residents. Tank-level sensing emerged as a low-cost, non-intrusive alternative to full smart metering, well suited to regions with intermittent supply. But the authors note a systemic blind spot: most water prototypes address a single function, and almost none consider the energy embedded in every liter pumped, heated, or treated.

That blind spot is the review’s central criticism. Households face coupled resource flows, where scheduling a water pump, a water heater, or a treatment load directly changes the electrical load profile, yet the overwhelming majority of research optimizes one resource in isolation. Multi-objective frameworks that jointly balance energy cost, water usage, occupant comfort, and equipment longevity remain underdeveloped. The authors also identify a gap between algorithms and reality: many models are trained on datasets from foreign households and may fail to generalize to homes with different occupancy patterns, climates, or appliance stocks. Economic evaluation is similarly thin, with studies reporting energy savings percentages but rarely installation costs, operating expenses, or payback periods. And while emerging directions such as edge AI, federated learning for privacy-preserving optimization, multi-agent reinforcement learning, transformer architectures, and hybrid swarm-intelligence optimizers promise more capable and adaptable systems, few have been tested in longitudinal, in-home deployments.

Security threads through every theme as a persistent brake on adoption. IoT smart homes capture intimate data streams, including motion patterns, appliance usage, and environmental conditions, and the reviewed studies repeatedly identified vulnerabilities in wireless channels, weak authentication, and unencrypted device-to-cloud transmission. Systems controlled through GSM networks or web dashboards are especially exposed to spoofing, interception, and unauthorized commands. Proposed countermeasures include graph convolutional networks for anomaly detection, deep learning-based security layers, device-level authentication, and encrypted communication, but the authors find that end-to-end security architectures, spanning secure transmission, access control, interoperability standards, and transparent user consent, are largely absent. Inexpensive control schemes built on platforms like WhatsApp demonstrate that smart home functionality can reach resource-limited settings, but they also illustrate how minimal the security layers in most prototypes truly are.

The review’s geographic framing gives it unusual urgency. Most studies originate in high-income countries with stable grids and mature digital infrastructure, while sub-Saharan Africa, where optimization could deliver the greatest impact, remains strikingly understudied. For Nigeria specifically, the authors see immediate openings: NILM could deliver appliance-level energy diagnostics without costly per-device metering; predictive and reinforcement learning controllers could synchronize appliance use with the country’s rapidly expanding rooftop solar capacity; and IoT tank monitoring and leak detection could reduce waste in a landscape dominated by boreholes and intermittent municipal supply. Their prescription is concrete: invest in local field deployments and multi-home datasets, build partnerships among universities, utilities, and technology firms, and design for resilience with edge computing, offline-capable systems, and low-cost hardware. The 32 studies synthesized here prove that AI can cut household energy use by double digits and watch over every drop of water. What remains is the harder engineering of integration, making one intelligent system manage both, affordably and securely, in the places that need it most.

Subject of Research: AI-based optimization of energy and water use in smart homes, based on a PRISMA systematic review of 32 studies from 2021 to 2025

Article Title: AI-driven smart home optimization for sustainable energy and water management: a systematic review

Article References: Ogbuanya, C. E., Eneh, J. N., & Udanor, C. N. (2026). AI-driven smart home optimization for sustainable energy and water management: a systematic review. Frontiers in Control Engineering, 7, Article 1793880. https://doi.org/10.3389/fcteg.2026.1793880

Image Credits: AI Generated

DOI: 10.3389/fcteg.2026.1793880

Keywords: smart homes, artificial intelligence, machine learning, reinforcement learning, home energy management systems, non-intrusive load monitoring, smart water management, Internet of Things, IoT security, systematic review, energy efficiency, Nigeria

The core problem the researchers tackled is twofold. First, conventional intrusion detection systems rely on enormous, carefully labeled datasets of network traffic. In IoT settings, such datasets are nearly impossible to assemble, because privacy concerns, thin monitoring infrastructure, and the sheer cost of manual annotation get in the way. Second, even the newer generation of few-shot learning systems, which are designed to learn from scarce data, turn out to be exquisitely vulnerable to adversarial examples: inputs that have been nudged by tiny, mathematically optimized perturbations designed to push them across a classifier’s decision boundary. A model can look brilliant on clean traffic and collapse catastrophically the moment an adversary bends it.

The team’s answer, which they call Adv-ProtoNet, fuses two ideas that had previously lived separate lives. The first is a Prototypical Network, a metric-based few-shot learning architecture. Instead of memorizing decision boundaries from millions of samples, the network learns an embedding function that maps network traffic into a space where each class of traffic, benign or malicious, is represented by a prototype, essentially the center of mass of its few labeled examples. A new packet is classified by measuring its squared Euclidean distance to each prototype and applying a softmax over those distances. The second idea is adversarial training, in which the model is deliberately attacked during training with perturbed versions of its own data so that it learns to withstand such attacks in the wild.

The integration is formulated as a min-max optimization problem. In the inner loop, the system generates the worst-case perturbation it can within a small, norm-bounded ball around each query sample, using one of four canonical gradient-based attacks: the Fast Gradient Sign Method (FGSM), the optimization-based Carlini and Wagner (C&W) attack, the boundary-seeking DeepFool, and the iterative Projected Gradient Descent (PGD), widely regarded as the gold-standard benchmark for first-order adversaries. In the outer loop, the network minimizes a weighted sum of the standard few-shot loss and the adversarial loss, with a hyperparameter lambda balancing clean accuracy against robustness. The effect, the authors explain, is to enforce local smoothness in the embedding space: small perturbations no longer drag a sample away from its true class prototype, so the entire neighborhood around each example remains inside the correct decision region.

Importantly, the researchers evaluated their system under a strict white-box threat model, the most pessimistic assumption possible. The adversary is assumed to know the model’s architecture, its feature representation, and its internal weights, and can therefore compute full gradients to craft highly optimized attacks. The security goals were correspondingly framed as worst-case guarantees against first-order adversaries. This is a deliberately conservative choice; a system that survives white-box attacks is far harder to defeat in the messier, information-poor conditions of a real network.

The experimental stage was the IDSIoT2024 dataset, a large and recent collection of more than 16 million network traffic samples spanning 96 features and seven classes: benign traffic plus six attack categories, including denial of service, injection, man-in-the-middle, malware, routing, and vulnerability analysis. The data was standardized to zero mean and unit variance, then split into 64 percent training, 16 percent validation, and 20 percent testing subsets with a fixed random seed to prevent leakage. Training followed an episodic protocol: across 5,000 episodes, each task sampled all seven classes with just three support examples per class to build the prototypes and five query examples to compute the loss. This three-shot regime deliberately mimics the data poverty of real IoT deployments.

Under benign conditions, the framework performed impressively. The baseline Prototypical Network achieved a mean accuracy of 93.83 percent, with precision at 93.72 percent and an F1-score of 93.75 percent. The class-wise picture was even more striking: the model correctly identified 99 percent of denial of service attacks and a perfect 100 percent of man-in-the-middle, malware, and routing attacks, with only the injection class lagging at 80 percent detection. For a system that never saw more than three labeled examples per class, this is a strong demonstration that metric learning can anchor reliable class prototypes even in extremely low-data regimes.

Then came the adversarial stress tests, and the initial results were sobering. Without adversarial training, accuracy collapsed to 11.43 percent under FGSM, 14.23 percent under C&W, 14.23 percent under DeepFool, and 14.46 percent under PGD, a near-random level of performance that shows how thoroughly perturbations scrambled the distance relationships in the embedding space. After integrating each attack into the episodic training loop, however, the picture transformed. Accuracy recovered to 86.57 percent against FGSM, 87.43 percent against C&W, 87.83 percent against PGD, and a remarkable 94.97 percent against DeepFool, effectively matching the clean baseline. The confusion matrices confirmed the recovery, with true-positive diagonals restored and class accuracies returning to 100 percent for several attack types under the DeepFool scenario.

Ablation studies and comparisons sharpened the message. A standard deep neural network, despite reaching 93.66 percent on clean data, fell to between roughly 43 and 51 percent under attack. A standard Prototypical Network did even better on clean data at 94.69 percent but collapsed to between 20 and 29 percent under adversarial pressure, proving that metric learning alone offers no immunity. Only the hybrid Adv-ProtoNet, trading a few points of clean accuracy for consistent robustness in the 78 to 83 percent range across all four attacks, held the line. Classical baselines constrained to the same three-shot regime fared far worse on clean data, with a support vector machine managing just 55.94 percent and a random forest 74.69 percent, compared with 95.71 percent for the standard ProtoNet. Matching Networks and Relation Networks, other state-of-the-art few-shot methods, matched or exceeded the clean accuracy of the proposed model but crumbled under PGD to below 21 percent, leaving Adv-ProtoNet as the only framework in the comparison that survived the full adversarial gauntlet.

Just as important for real-world deployment is the computational profile. The framework’s memory footprint stays between roughly 20 and 20.4 megabytes even with adversarial optimization enabled, because the perturbation generation reuses existing computational graph allocations. Meta-training, which takes between about 32 and 182 seconds depending on the attack, runs offline on centralized hardware, so the burden never touches the IoT edge. Inference is the critical metric for live defense, and here the numbers are compelling: standard queries are classified in 0.06 to 0.09 milliseconds, and even the heaviest adversarial evaluation stays under one millisecond. The authors caution that future work must extend the approach to black-box and transfer-based attacks and to lightweight adaptations for real-time edge deployment, but the study makes a persuasive case that the dual challenges of data scarcity and adversarial vulnerability in IoT security are not separate problems to be solved in isolation. They are, it turns out, best solved together, by teaching a model to recognize attacks from almost nothing and to hold its ground when those attacks fight back.

Subject of Research: Adversarially robust few-shot learning for intrusion detection in Internet of Things networks

Article Title: Adversarially robust few-shot learning for aI-enabled attack mitigation in the internet of things

Article References: Asiri, F., Alturki, N., Alnazzawi, N., Khattak, A. A., Ali, H., Naeem, U., & Latif, S. (2026). Adversarially robust few-shot learning for aI-enabled attack mitigation in the internet of things. Cluster Computing, 29(12), Article 731. https://doi.org/10.1007/s10586-026-06528-5

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06528-5

Keywords: Internet of Things, few-shot learning, adversarial training, intrusion detection, prototypical networks, cybersecurity, machine learning, adversarial attacks, FGSM, PGD, network security, resource-constrained devices