Privacy-Preserving AI Models Using Homomorphic Encryption in Federated Learning Environments

The increasing reliance on distributed artificial intelligence (AI) systems has raised significant concerns regarding data privacy and security, particularly in sensitive domains such as healthcare and finance. Federated learning (FL) has emerged as a promising paradigm for decentralized model training by allowing data to remain at its source while sharing model updates. However, traditional FL frameworks are still vulnerable to information leakage during communication and aggregation processes. This study proposes a privacy-preserving AI framework that integrates homomorphic encryption (HE) into federated learning environments to enhance data security while maintaining predictive performance. The performance analysis demonstrates that privacy-preserving mechanisms introduce a measurable trade-off between model accuracy and security. The baseline model without encryption achieves the highest accuracy, while secure aggregation results in a slight reduction. Homomorphic encryption, providing the strongest privacy guarantees, introduces a modest decrease in accuracy due to computational constraints. Despite this reduction, the performance remains within acceptable limits, indicating the feasibility of HE-based approaches in practical applications. In addition to accuracy, the study evaluates computational overhead associated with privacy-preserving techniques. The results show that homomorphic encryption significantly increases processing time per training round compared to unencrypted models, highlighting the need for optimization strategies. However, the enhanced security benefits justify this overhead in scenarios requiring strict data protection. Furthermore, the analysis of privacy–performance trade-offs reveals that increasing privacy levels leads to gradual declines in model accuracy. This finding underscores the importance of balancing security requirements with predictive performance when designing AI systems.