Lightweight Client Weighting for Robust Federated Learning Under Label-Flipping Attacks

Federated learning (FL) enables privacy-preserving machine learning across decentralized clients without sharing raw data. However, it faces significant challenges, particularly from label-flipping (LF) attacks, where malicious clients mislabel data, compromising model performance. This paper proposes a robust aggregation method that mitigates the impact of LF attacks by combining validation-based weighting with update consistency. This method adaptively adjusts client weights based on their validation accuracy and the consistency of their model updates across rounds. Unlike traditional approaches, which rely on costly computation or access to trusted validation sets, this lightweight method requires no access to local data or extensive computation, making it suitable for real-world applications. Experimental results using the Fashion-MNIST dataset show that the proposed method effectively maintains competitive accuracy under LF attacks, outperforming standard aggregation methods such as FedAvg and Trimmed Mean. This strategy offers a practical, scalable solution for FL in adversarial settings, providing robust defense against label-flipping attacks without compromising the system's efficiency.