Siamese Network Based Anomaly Detection Framework for Robust Federated Learning

Federated learning is a decentralized approach to machine learning that has increased in popularity in recent years. It enables several participants to train a common model without revealing their data. This strategy is, however, susceptible to attacks from malicious clients who can launch targeted model poisoning attacks and reduce learning performance by delivering false model updates to the server. It is necessary to identify and eliminate such fraudulent updates and the attackers behind them to preserve the robustness and security of the shared model. In this paper, a novel Siamese network-based architecture for robust federated learning is proposed. which can identify and eliminate harmful updates. Our method is assessed and compared with other approaches for adversarial detection in image classification tasks in a federated setting, using a CNN model. Experimental findings demonstrate that the system offers reliable federated learning that is resistant to both targeted model poisoning and untargeted Byzantine attacks.