Aug 2026· Italian National Conference on Sensors· Vol 26· 0 citations· 99 references
Medicine
TL;DR
Future research should emphasize feasible attack paradigms, hierarchical lightweight detection, reliable multimodal fusion, certifiable operational boundaries, and auditable end-to-end response mechanisms, thereby supporting the evaluation and deployment of adversarial example detection in IoT scenarios.
Abstract
Deep learning has been widely deployed in critical scenarios such as the Internet of Things (IoT), industrial sensing, network intrusion detection, and cyber-physical system monitoring, where model inference directly affects system security, operational reliability, and service continuity. However, existing adversarial example detection studies remain insufficient for practical IoT deployment, as their validation often overlooks endpoint resource constraints, heterogeneous data modalities, physical environmental interference, communication protocol specifications, adaptive attacks, and adversary capability models. Moreover, detection outcomes are rarely connected with deployment locations, computational overhead, formal security assurance, and subsequent response strategies, which limits their engineering applicability. To address these limitations, this review systematically synthesizes recent representative studies in adversarial example detection and constructs a unified analytical framework integrating detection evidence, IoT deployment feasibility, and adaptive-attack evaluation. Based on the source of detection evidence, existing methods are categorized into input-consistency-based, feature-statistics-based, predictive-uncertainty-based, model-reconstruction-based, runtime-context-aware, and multi-strategy fusion detection, while formal certification is discussed as an independent security-assurance dimension. The review further analyzes the principles, applicable conditions, limitations, compatibility conflicts with IoT deployment constraints, and typical failure modes of these methods. The analysis identifies four key challenges: the lack of IoT-native adaptive evaluation, limited anomaly-boundary identification and cross-modal generalization, insufficient deployment-time security assurance, and weak coordination between detection decisions and security responses. Future research should therefore emphasize feasible attack paradigms, hierarchical lightweight detection, reliable multimodal fusion, certifiable operational boundaries, and auditable end-to-end response mechanisms, thereby supporting the evaluation and deployment of adversarial example detection in IoT scenarios.
The rapid proliferation of Internet of Things (IoT) devices across critical domains including healthcare, smart cities, industrial control systems, and intelligent transportation has fundamentally transformed the cybersecurity threat landscape. The inherent characteristics of IoT environments, namely resource-constrained devices, heterogeneous architectures, and large-scale deployment, render traditional Intrusion Detection Systems (IDS) inadequate for the sophisticated and evolving attack vectors targeting these networks. Deep learning (DL) has emerged as a compelling paradigm for next-generation IoT IDS, offering automated feature extraction, temporal pattern recognition, and adaptive threat detection capabilities that address the limitations of conventional approaches.
This paper provides a thorough and systematic review of the existing DL methods for IoT intrusion detection. The paper explore the IoT architectural paradigms, outline a four layered taxonomy for types of IoT attacks across its three primary layers Perception, Network and Application as well as Adversarial Machine Learning attacks, and systematically review seven classes of DL architectures Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), Autoencoders, Generative Adversarial Networks (GAN), models based on Transformer architecture and Federated Learning frameworks. In a comparative review of forty peer-reviewed studies, we demonstrate that hybrid DL models provide excellent detection performance (99-100% classification accuracy on benchmark datasets) as well as practical viability for deployment with privacy-preserving Federated Learning for large-scale data. The study additionally highlights five enduring challenges class imbalance, adversarial vulnerability, zero-day detection limitations, computational constraints and the absence of standardized benchmarking protocols that together account for the gap between performance benchmarks and real-world deployment efficacy. It outlines future research avenues targeting on five key axes with a particular focus in the integration of Explainable AI (XAI), lightweight edge-deployable architectures, and adversarial robustness mechanisms. This survey identifies a structured reference to advance the state of IoT intrusion detection from research to operationally viable and deployable systems.
Mohammed Gharkan, Mustafa I. Hussien Al-Janabi, Obaid Salim· Al-Noor Journal of Engineeri...· 0 citations
This SLR analyzes 50 research articles to provide a holistic view of AML threats in IIoT systems and identifies seven distinct attack types: gradient-based perturbations, GAN-generated samples, poisoning attacks, reinforcement learning-based (RL) strategies, saliency-based feature manipulation, false data injection, and hybrid approaches.
Abeer Alsaidlani, Muhammad Rashid, M. Aljabri· Italian National Conference...· 0 citations
A constraint-aware adversarially robust Internet of Things (IoT) traffic classification system with protocol validity, device behavior consistency, and manifold-aware training and evaluation is presented, demonstrating improved robustness, realism, and deployability compared to existing approaches.
With the proliferation of Internet of Things (IoT), critical infrastructures such as smart cities, industrial automation, precision healthcare, and intelligent transportation have seen a significant shift in their security landscape. As these critical infrastructures are increasingly depending on OneM2M standards for interoperability and scalable service management, they are simultaneously presenting a larger and more structured attack surface. One of the most serious and realistic threats in such scenarios is black-box adversarial attacks, in which an attacker can compromise Machine Learning (ML) and Deep Learning (DL)-based intrusion detection systems (IDS) without having prior knowledge of the ML/DL model's architecture, parameters, and training data. To mitigate such a critical challenge, we propose a multi-stage framework specific to OneM2M-based IoT networks that can effectively counter transfer-based and query-based black-box adversarial attacks. Our framework's effectiveness will be tested on a dataset of 1.25 million labeled network flows collected from Azure IoT Hub-based IoT devices over a 10-day period. We will test the robustness of our framework against three adversarial attacks: FGSM with $\varepsilon=0.05$, PGD with 40 iterations and step size 0.01, and C&W with L2 optimization and 1,000 iterations. The proposed pipeline consists of four different mechanisms: (1) adversarial training with the use of mixed clean and adversarial samples, (2) traffic sanitization to minimize malicious and/or suspicious traffic flows before processing by the Common Service Entity (CSE), (3) the use of a black-box adversarial detector based on feature transformation, and (4) the use of an ensemble-based IDS to make use of the results of the Random Forest, XGBoost, MLP, and 1D-CNN-based IDSs via majority voting. The experimental results show that the proposed Ensemble + multi-defense configuration results in 98.1% accuracy, 98.3% TPR, and 1.7% FNR. Most notably, the proposed Ensemble + multi-defense configuration results in an ASR of 11.2%, which is down by 86% compared to the 82.3% ASR achieved by the baseline IDS. Notably, the proposed framework results in limited overheads in terms of RTT (increased by 8 ms), CPU (increased by 17%), and RAM (increased by 1.6 GB).
Hamza Jamiri, A. Zyane· International Conference on...· 0 citations
Introduction: The rapid expansion of Internet of Things (IoT) deployments has increased the cyber-attack surface and introduced heterogeneous traffic behaviour across devices, gateways, edge services, and network environments. Although many IoT intrusion detection studies report high performance under independent and identically distributed test conditions, such results often provide limited evidence of real deployment reliability, particularly for unseen hosts, cross-dataset transfers, and calibration drifts.
Methodology: This study presents a deployment-aware evaluation and calibration framework for IoT intrusion detection using a Denoising Autoencoder-Based Deep Feature Extraction (DAE-DFE) backbone. Rather than proposing a new neural architecture, this study focuses on robustness-oriented evaluation protocols and calibration-aware decision-making under domain shifts. The framework was evaluated using conventional IID splits, identifier-removed testing, and source-IP-based GroupSplit evaluation on NF-ToN-IoT-v2 to reduce the memorisation of the host. Cross-dataset robustness was assessed by converting Edge-IIoTset packet/protocol logs into pseudo-flows and testing Edge→NF and NF→Edge transfer using unsupervised threshold adaptation based on positive rate matching.
Results: On the NF-ToN-IoT-v2 GroupSplit, the framework achieved an F1 score of 0.9919 and ROC-AUC of 0.9997. In the Edge→NF cross-dataset setting, the model retained a meaningful ranking performance with ROC-AUC = 0.7962, while unsupervised threshold adaptation improved the target-domain accuracy from 0.4959 to 0.8878 and F1 score from 0.6630 to 0.8981.
Conclusion: The findings show that calibration-aware thresholding and deployment-realistic evaluation are essential for assessing IoT IDS reliability beyond the conventional IID accuracy.
Ola Madi Mohammed Al Mari· Majestic International Journ...· 0 citations
This survey research work proposes a lifecycle-based understanding of AI security threats and proposed a unified taxonomy for the four major categories of threats observed in real-world settings, namely, data poisoning and backdoor attacks on learning model updates, adversarial attacks on model outputs through input manipulation, privacy leakage through model-based queries, and model extraction for intellectual property theft and creation of rogue replicas of learning models.
Venkatesan Cherappa, Hsin-Hung Cho, Yasir Abdullah Rabi et al.· Journal of Internet Technolo...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
