Optimizing Privacy-Accuracy Trade-Off in IoT Intrusion Detection: An Analysis of FedAvg and FedProx with Differential Privacy

Ensuring the privacy and accuracy of interconnected network devices is crucial in the Internet of Things (IoT) landscape, especially when deploying Intrusion Detection Systems (IDS) on devices with constrained resources. This study offers a method to improve intrusion detection in IoT environments by combining Federated Learning with Differential Privacy. The proposed approach builds an efficient intrusion detection model by fusing Bidirectional Long Short-Term Memory (BiLSTM) with Convolutional Neural Networks (CNN). We evaluate this model on the Bot-IoT dataset, meticulously curated by the University of New South Wales (UNSW). This work aims to balance high detection accuracy with robust privacy protection, a crucial yet often neglected aspect of IoT security. We use federated learning to distribute intrusion detection tasks across various IoT devices while protecting data privacy, using differential privacy algorithms to assess and limit information leakage. This study evaluates two aggregation strategies, Federated Averaging (FedAvg) and Federated Proximal (FedProx), and examines how varying epsilon values and noise multipliers affect model accuracy. The findings indicate t hat a noise multiplier o f roughly 0.5 yields a steady accuracy curve, with FedProx achieving higher accuracy than FedAvg.