ABSTRACT The escalating deployment of Internet of Things (IoT) devices across critical infrastructure demands intrusion detection systems (IDS) capable of reliably discriminating hierarchically structured cyber threats such as denial‐of‐service, man‐in‐the‐middle, Mirai, and scan attacks. This study proposes a hybrid CNN‐GRU‐attention architecture in which parallel convolutional and recurrent branches are fused through a cross‐modal attention mechanism computing joint scores from spatial and temporal features via a shared scoring function, enabling adaptive per‐sample prioritisation of the more discriminative branch. ADASYN and ProWRAS oversampling strategies are compared within a leakage‐free pipeline across binary, category and sub‐category classification on IoTID20, and validated independently on TON‐IoT against six deep learning baselines. On IoTID20, test accuracies reach 1.0000, 0.9995 and 0.8228/0.8225 (ADASYN/ProWRAS). McNemar's tests show no significant difference between samplers, although per‐class analysis reveals a precision–recall trade‐off at the sub‐category tier. On TON‐IoT, the model achieves 0.9788 accuracy and F1‐score, exceeding six baselines on most metrics, trailing only in balanced accuracy due to one underrepresented class. Attention‐weight analysis shows concentration decreasing with granularity, with ProWRAS inducing a more distributed pattern that explains this trade‐off. An ablation study confirms attention and oversampling chiefly improve precision (0.8019–0.9005) over accuracy, though residual Mirai confusion reflects a feature‐space limitation.
ABSTRACT The rapid growth of Internet of things (IoT) deployments has intensified the need for effective and reliable intrusion detection systems capable of operating under heterogeneous and resource‐constrained environments. In response, deep learning (DL)‐based intrusion detection systems have been increasingly explored across IoT application domains. This study presents a PRISMA guided systematic review with descriptive meta‐analysis of recent DL‐based intrusion detection research in IoT environments, focusing on architectural trends, evaluation practices, deployment considerations and trust‐related properties. A total of 42 primary empirical studies published between 2021 and early 2026 were identified through systematic searches across major scholarly databases and analysed using standardized inclusion criteria and data extraction protocols. Instead of statistically aggregating heterogeneous results, the review employs descriptive meta‐analytic synthesis to examine patterns in model architectures, learning paradigms, deployment orientations and evaluation metrics. The analysis reveals a gradual shift from centralised DL models towards hybrid architectures, federated learning frameworks and resource‐aware designs suitable for edge or fog environments. Federated and privacy‐preserving learning approaches demonstrate promising detection performance while maintaining data locality, whereas explainable AI techniques are incorporated in only a limited subset of current IoT intrusion detection systems. The results also highlight a persistent gap between benchmark‐driven evaluation practices and the operational requirements of IoT deployments. Future studies should therefore adopt evaluation practices that jointly consider detection effectiveness, computational cost, privacy preservation, interpretability and deployment feasibility.
ABSTRACT Generative black‐box fuzzing techniques have been demonstrated to offer distinct advantages when confronted with closed‐source systems. However, testers have been shown to lack efficiency in developing fuzzing templates for emerging increasingly complex protocols. To address this challenge, we develop GeniFuzz and construct a proprietary knowledge database for generative fuzzing. This database is designed to refine and contextualise the large language models, thereby enabling it to substitute for the experts' function in generating fuzzing templates. Thus, a generative fuzzing framework based on large language models has been established. The experimental results demonstrate that GeniFuzz's composite score in the fuzzing template generation task is enhanced by 37.41% on average compared to the expert‐written templates, and by 22.48% over the expert‐written templates in terms of path coverage, and by 17.22% on average in efficiency of triggering crashes. Furthermore, the experiments demonstrate that the enhancement in the efficacy of fuzz testing, as implemented by the GeniFuzz framework, exhibits robustness and independence from the particular large language model utilised.
ABSTRACT This paper proposes a collision‐aware proximal policy optimisation (CA‐PPO) framework for adaptive resource management in grant‐free nonorthogonal multiple access (GF‐NOMA) enabled wireless federated learning systems. Unlike conventional approaches that rely on static clustering or grant‐based access, CA‐PPO dynamically coordinates device clustering and transmit‐power allocation under stochastic channel conditions, random‐access collisions and successive interference cancelation (SIC) constraints. The proposed framework employs a hybrid discrete–continuous actor–critic architecture with a collision‐aware reward that jointly captures decoding success, energy consumption, latency and retransmissions. CA‐PPO enables scalable and reliable aggregation in ultradense IoT deployments by improving participation stability and reducing communication overhead. Simulation results demonstrate that the proposed approach accelerates convergence by approximately 16%, reduces per‐device energy consumption by up to 25% and sustains high device participation under dense connectivity. These findings highlight the potential of learning‐driven network‐aware resource optimisation for next‐generation wireless federated learning systems.
ABSTRACT In modern computing and blockchain systems, with the growth of networks, classical consensus mechanisms, which tend to be hard and rule‐based, cannot dynamically adapt to changing workloads and, therefore, consume unnecessary energy and have uneven performance. The proposed AI–blockchain convergence framework is aimed at accomplishing energy‐efficient and self‐optimising computing networks based on reinforcement‐based consensus adaptation. The proposed model combines a multi‐agent reinforcement learning controller with blockchain telemetry which allows the real‐time adjustment of the block size, committee composition, and timeouts. Compared to the traditional consensus system, AIBLOCK adapts to both throughput and stability based on feedback of the live states on a network, inherently turning the blockchain into a self‐regulating digital organism. Large scale node deployments have been evaluated through experimentation and have a 34% lower energy use as well as a 22% increase in consensus stability when compared to HiCoOB and Layer‐2 baselines. The framework reported sublinear scaling of energy per node, and it converges quickly when subjected to dynamic loads in terms of transactions, which validate its adaptive efficiency. AIBLOCK, therefore, represents a transition to cognitive blockchain ecosystems that do not necessarily have to be manually tuned but learn instead.
