PerFedHypID: A Personalized Federated Hypernetworks based aggregation approach for Intrusion Detection Systems

Abstract Traditional Network Intrusion Detection Systems (NIDS) face scalability challenges due to the vast amount of data generated by Internet of Things (IoT) devices, coupled with growing privacy concerns. Federated Learning (FL) emerges as a distributed, privacy-preserving learning paradigm that trains deep learning models locally, mitigating privacy risks associated with centralized data processing. However, conventional FL strategies entail limitations. First, they require all clients to use the same model architectures. This makes personalized learning harder, especially when there is non-IID (independent and identically distributed) heterogeneous data. Secondly, the weight aggregation process in FL introduces communication overhead, potentially slowing down aggregation. Also, even though there are encryption methods like homomorphic encryption and differential privacy, encrypted weights take more time to process and can be hacked to see how the data is distributed. This is especially true in dynamic IoT environments where attack types cause data distributions to change all the time. To address these challenges, we propose PerFedHypID: A Personalized Federated Hypernetworks-based aggregation strategy for Intrusion Detection Systems. Our approach incorporates embedding vectors instead of weights, which are lighter and offer enhanced personalization capabilities. PerFedHypID leverages personalized layers and hypernetwork aggregation for efficient and personalized model aggregation. We extensively evaluated our proposed system on the CSE-CICIDS-2018 dataset under various non-IID heterogeneous settings. Our results demonstrate robust performance compared to state-of-the-art personalized federated learning algorithms.