C-PFL: A committee-based personalized federated learning framework

Federated Learning (FL) is an emerging machine learning paradigm that enables multiple parties to train a shared model while preserving data privacy collaboratively. However, malicious clients pose a significant threat to FL systems. This interference not only deteriorates model performance but also exacerbates the unfairness of the global model caused by data heterogeneity, leading to inconsistent performance across clients. We propose C-PFL, a committee-based personalized FL framework that improves both robustness and personalization. In contrast to prior approaches such as FedProto (which relies on the exchange of class prototypes), Ditto (which employs regularization between global and local models), and FedBABU (which freezes the classifier head during federated training), C-PFL introduces two principal innovations. C-PFL adopts a split-model design, updating only a shared backbone during global training while fine-tuning a personalized head locally. A dynamic committee of high-contribution clients validates submitted updates without public data, filtering low-quality or adversarial contributions before aggregation. Experiments on MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and AGNews show that C-PFL outperforms six state-of-the-art personalized FL baselines by up to 2.89% in non-adversarial settings, and by as much as 6.96% under 40% malicious clients. These results demonstrate C-PFL’s ability to sustain high accuracy and stability across diverse non-IID scenarios, even with significant adversarial participation.