Improving adversarial resilience for anomaly detection in the heterogeneous internet of things through ensemble models

초록

With the continuous expansion and increasing complexity of the Internet of Things (IoT), anomaly detection systems have become prime targets for sophisticated adversarial attacks. These attacks often exploit weaknesses in existing detection frameworks, particularly under conditions of class imbalance and dynamic, heterogeneous data streams. To address this challenge, we propose a robust and scalable ensemble deep learning framework that integrates Conditional Generative Adversarial Networks (cGANs), Denoising Autoencoders (DAEs), and Long Short-Term Memory (LSTM) networks for anomaly detection in IoT environments. Specifically, the framework leverages cGANs to synthesize minority-class samples and alleviate data imbalance, employs DAEs for robust and noise-resilient feature extraction, and utilizes LSTM networks to capture temporal dependencies inherent in sequential IoT data. To further enhance resilience against evasion attacks, we incorporate a tailored multi-layer adversarial training strategy using both clean and dynamically generated adversarial samples along with partial gradient masking. In addition, we introduce a lightweight knowledge distillation framework, enabling a compressed student model to achieve comparable accuracy with reduced inference delay, thereby improving deployment feasibility on edge devices. Our contributions are fivefold: (i) we develop a novel ensemble architecture designed for robust and resilient anomaly detection in heterogeneous IoT systems; (ii) we introduce a customized adversarial training approach optimized for real-time constraints in IoT settings; (iii) we implement a lightweight feature selection and distillation pipeline for complexity reduction; (iv) we conduct comprehensive evaluations using the Distributed Smart Space Orchestration System (DS2OS) and Bot-IoT datasets, achieving strong performance across domains (F1: 96.26% on DS2OS, 95.94% on Bot-IoT).; and (v) we demonstrate that the proposed framework consistently outperforms state-of-the-art standalone and hybrid methods across a range of attack scenarios. Overall, the proposed system offers a practical and scalable defense mechanism against emerging threats in future IoT infrastructures.

키워드