Towards Secure Federated Learning: The Impact of Cost-Sensitive Learning on Adversarial Resilience

Real-world Federated learning (FL) applications often suffer from imbalanced data distributions, where some classes are significantly underrepresented, leading to poor generalization on minority classes. This imbalance poses a critical challenge, especially in high-stakes domains like healthcare and fraud detection, where detecting rare events is essential. Moreover, FL systems are vulnerable to adversarial attacks that exploit these weaknesses by manipulating model updates, further degrading performance. To address these challenges, we propose to integrate cost-sensitive learning into the FL process, using class weighting to adjust the loss function, thereby enhancing the model’s ability to learn from underrepresented classes. Additionally, we simulate various adversarial attacks, including parameter manipulation and noise injection, to evaluate the robustness of the model. We employ parameter clipping as a defense strategy to mitigate the impact of these attacks on the global model. Our preliminary experiments suggest that cost-sensitive learning may improve accuracy for minority classes without notably degrading overall model performance. Additionally, applying parameter clipping appears to strengthen the model’s resistance to adversarial manipulations, potentially contributing to a more robust and equitable federated learning system. This work aims to develop a library that addresses the challenges of data imbalance and adversarial robustness in federated learning, with the goal of providing a practical tool for real-world deployments.