A Summary of Adversarial Attacks and Defenses on ML- and Hardware-based IoT Device Fingerprinting and Identification [Póster]

Abstract

In response to the rapid expansion of Internet of-Things (IoT) devices and associated cybersecurity threats, this work proposes a novel LSTM-CNN architecture for robust individual device identification, leveraging behavior monitoring and ML/DL advancements. Evaluated against a dataset from 45 Raspberry Pi devices, this model outperforms traditional ML/DL methods, achieving a +0.96 average F1-Score and demonstrating strong resilience to adversarial attacks, including context-based and ML/DL-specific evasion attempts. Through the application of adversarial training and model distillation defenses, the model vulnerability to the most effective attack was reduced from a 0.88 success rate to 0.17, maintaining high-performance integrity.