Integrating Differential Privacy and HE for Enhanced Privacy and Security in FL without Compromising Model Utility

Training machine learning models in a federated fashion across numerous devices or servers while maintaining data locality is an exciting new paradigm. However, maintaining data confidentiality throughout the aggregating process is still difficult. By combining DP with HE, this research presents a fresh strategy for implementing privacy-protecting FL. To avoid any accidental disclosure of private information, our technique keeps the data encrypted throughout the training process. Using DP in the big model keeps it safe from sharing details about any single data point. This is done by adding random signals into how updates are made to the model. The balance between making sure privacy is safe, the amount of work needed and how accurate a model can be are discussed here. The results of the tests show that our method works better than others and gives good model performance with strong protection for privacy. This study paves the way for a new time of private and secure FL. In this era, user privacy is taken into account without affecting how well these models work at all.