Efficient Techniques for Differential Privacy in Deep Learning
Shreya G. Tendulkar · 2025
Privacy leakage in deep learning models trained on sensitive data is a growing concern, particularly in preventing data regurgitation and membership inference attacks. Differential Privacy (DP) provides a formal framework to mitigate these risks, yet conventional approaches like Differentially Private Stochastic Gradient Descent (DP-SGD) suffer from substantial computational overhead and performance degradation. This work introduces novel techniques to improve DP training efficiency, particularly for large pretrained models such as Transformers. Empirical results show that strategic optimizations can significantly enhance the privacy-utility trade-off while reducing the high memory demands associated with DP implementations. Furthermore, theoretical analyses challenge existing assumptions, demonstrating dimension-independent performance bounds in specific optimization scenarios. These findings contribute to advancing differentially private machine learning, with practical implications in industry, including integration into Microsoft's privacy-preserving initiatives. The research underscores the importance of optimizing DP techniques for real-world applications and provides a foundation for further innovations in safeguarding sensitive data.