Detection and Mitigation of Adversarial Attacks in Federated Learning for IoT Networks

Abstract

The rapid expansion of the Internet of Things (IoT) has transformed data collection across critical sectors, particularly healthcare. While Federated Learning (FL) offers a promising paradigm by enabling collaborative model training without centralizing sensitive data, its distributed architecture introduces significant security and privacy vulnerabilities. This thesis investigates the detection and mitigation of adversarial attacks in FL systems deployed over IoT networks. We focus on a realistic cross-silo medical scenario in which multiple hospitals collaboratively train a chest X-ray disease detection model based on the DenseNet-121 architecture. We first demonstrate the vulnerability of standard FL to white-box Membership Inference Attacks (MIA), revealing how data overlap between training and test sets creates exploitable structural backdoors that enable an attacker to determine whether a specific patient’s data was used during training. To counter these threats, we propose a progressive defense-in-depth strategy combining three complementary layers: (1) structural data reorganization through patient-level splitting combined with Low-Rank Adaptation (LoRA) fine-tuning, (2) Selective Differential Privacy with temperature scaling for mathematical noise injection, and (3) Homomorphic Encryption (HE) for securing the communication channel between clients and the server. Experimental results demonstrate that our integrated layered defense effectively reduces the attacker’s success rate to random chance while preserving the diagnostic performance of the model, proving that secure and privacy-preserving collaborative AI over IoT networks is practically achievable.