Differentially Private Clipped-SGD: High-Probability Convergence with Arbitrary Clipping Level
This work addresses a gap in optimization theory for differentially private deep learning, enabling better trade-offs between convergence speed and privacy guarantees, though it is incremental as it builds on existing clipping and DP methods.
The paper tackles the incompatibility between high-probability convergence analyses requiring increasing clipping thresholds and fixed clipping levels needed for differential privacy mechanisms, showing that DP-Clipped-SGD with a fixed clipping level converges to a neighborhood of the optimal solution with a faster rate than existing methods under heavy-tailed noise.
Gradient clipping is a fundamental tool in Deep Learning, improving the high-probability convergence of stochastic first-order methods like SGD, AdaGrad, and Adam under heavy-tailed noise, which is common in training large language models. It is also a crucial component of Differential Privacy (DP) mechanisms. However, existing high-probability convergence analyses typically require the clipping threshold to increase with the number of optimization steps, which is incompatible with standard DP mechanisms like the Gaussian mechanism. In this work, we close this gap by providing the first high-probability convergence analysis for DP-Clipped-SGD with a fixed clipping level, applicable to both convex and non-convex smooth optimization under heavy-tailed noise, characterized by a bounded central $α$-th moment assumption, $α\in (1,2]$. Our results show that, with a fixed clipping level, the method converges to a neighborhood of the optimal solution with a faster rate than the existing ones. The neighborhood can be balanced against the noise introduced by DP, providing a refined trade-off between convergence speed and privacy guarantees.