Adversarial Threats and Defence Mechanisms in Artificial Intelligence of Things Systems: A Systematic Review

초록

Artificial Intelligence of Things (AIoT) systems have emerged through the rapid integration of artificial intelligence (AI) and the Internet of Things (IoT), enabling intelligent sensing, distributed learning, and realtime decision-making across diverse application domains. However, this convergence also introduces a significantly expanded adversarial attack surface spanning sensing devices, communication networks, learning pipelines, and actuation environments. This paper presents a comprehensive systematic review of adversarial threats and defence mechanisms in AIoT systems using a novel 3D-AIoT-TT (Three-Dimensional AIoT Threat Taxonomy) framework. The proposed taxonomy jointly models three fundamental dimensions: (i) AI pipeline stages, (ii) IoT architectural layers, and (iii) adversarial knowledge levels. By integrating these dimensions into a unified analytical framework, the taxonomy enables systematic characterization, classification, and evaluation of adversarial behaviours across heterogeneous AIoT environments. Using this framework, the study analyzes a broad spectrum of adversarial threats, including poisoning attacks, evasion attacks, backdoor attacks, model extraction attacks, and federated learning-based attacks. In addition, existing defence mechanisms such as adversarial training, hardware-assisted protection, federated defence strategies, anomaly detection, and certified robustness techniques are critically examined with particular emphasis on their practicality in resource-constrained edge AIoT deployments. A key contribution of this work is the identification and analysis of emergent cross-layer adversarial threats, where vulnerabilities arise through interactions among sensing, communication, learning, and actuation components. Unlike traditional isolated security models, these threats propagate across multiple AIoT layers and exploit systemic interdependencies within intelligent infrastructures. Overall, this study establishes a structured and unified understanding of adversarial AIoT security, highlights current research limitations and underexplored areas, and emphasizes the need for scalable, cross-layer, and deployment-aware defence mechanisms for next

키워드