A Robust Horizontal Federated Learning Framework Based on Trimmed K-Means and Homomorphic Encryption

To address the challenges of model degradation and security threats caused by data heterogeneity and poisoning attacks in Horizontal Federated Learning, this paper proposes a robust and privacy-preserving horizontal federated learning framework based on Trimmed K-Means clustering and Homomorphic Encryption. The proposed method replaces traditional K-Means with a Trimmed K-Means algorithm to achieve robust local outlier detection, mitigating the distortion of cluster centers caused by malicious data. Each client performs local anomaly filtering and applies the Paillier Homomorphic Encryption scheme to encrypt its gradient updates, ensuring privacy preservation during communication and aggregation. On the server side, a Dynamic Trust Evaluation Mechanism is introduced to continuously assess the reliability of participating clients through temporal trust evolution and cluster consistency analysis. The framework further employs robust aggregation algorithms to replace conventional averaging, thereby enhancing resilience against coordinated poisoning and adversarial updates. Experimental results demonstrate that the proposed framework maintains over 90% detection accuracy and less than 3% performance degradation even when up to 40% of clients are malicious, significantly outperforming traditional clustering-based evaluation schemes. This work effectively enhances the robustness, security, and trustworthiness of federated learning systems under adversarial and privacy-sensitive environments.