Resilience-Aware Quarantine Selection Mechanism for Federated Learning Environments

Federated Learning (FL) provides a promising approach for training Machine Learning (ML) models in a decentralized manner, ensuring the security of the global model while safeguarding participant privacy. In a typical FL scenario, clients often exhibit heterogeneous data distributions and varying hardware configurations. Additionally, the server must rely on the assumption that clients have correctly updated their models to the edge server. However, malicious clients can damage the global model by sending and aggregating harmful updates. Therefore, client selection algorithms must address the challenges of heterogeneous data distribution and the presence of malicious clients to ensure resilience against poisoned updates, while also enhancing the accuracy and integrity of the FL system. This paper introduces RiQS, a resilient and robust client selection algorithm designed to protect the global model from malicious attacks using a quarantine and similarity detection approach. RiQS employs Centroid-Based Kernel Alignment (CKA) to assess the similarity of the selected clients' models. This similarity information is then used to identify and exclude malicious clients from contributing to the edge server's model aggregation. RiQS maintains an accuracy of approximately 90% across all training rounds, effectively countering three types of malicious attacks: zero, random, and shuffle parameter attacks.