PLGuard: A Robust Aggregation Framework Against Poisoning Attacks in Federated Learning

Federated learning, a distributed machine learning paradigm, enables collaborative modeling while preserving data privacy. However, this framework faces severe security threats from malicious participants who upload poisoned model updates to corrupt the global model. To address this challenge, this paper proposes PLGuard, a robust aggregation framework based on comprehensive pseudo-label credibility assessment. The core innovation involves using a small public validation set to evaluate client credibility through multi-dimensional analysis of pseudo-labels generated by local models, considering their accuracy, calibration, and distributional consistency across training rounds. The server dynamically updates trust scores and performs adaptive weighted aggregation to suppress malicious updates. Extensive experiments on Fashion-MNIST, CIFAR-10, and MNIST datasets demonstrate that PLGuard significantly enhances global model accuracy and greatly reduces the success rates of label-flipping, backdoor, and model poisoning attacks, while maintaining stable convergence in heterogeneous environments. The results verify that PLGuard can effectively distinguish between benign and malicious clients, providing a practical and efficient defense framework for real-world federated learning systems under diverse adversarial threats.