On the Convergence and Calibration of Deep Learning with Differential Privacy
arXiv:2106.07830
Abstract
Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the lens of neural tangent kernel (NTK), which characterizes the per-sample gradient clipping and the noise addition in DP training, for arbitrary network architectures and loss functions. Interestingly, we show that the noise addition only affects the privacy risk but not the convergence or calibration, whereas the per-sample gradient clipping (under both flat and layerwise clipping styles) only affects the convergence and calibration. Furthermore, we observe that while DP models trained with small clipping norm usually achieve the best accurate, but are poorly calibrated and thus unreliable. In sharp contrast, DP models trained with large clipping norm enjoy the same privacy guarantee and similar accuracy, but are significantly more \textit{calibrated}. Our code can be found at \url{https://github.com/woodyx218/opacus_global_clipping}.
References in corpus (8)
- On Calibration of Modern Neural Networks
- Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks
- Gaussian Differential Privacy
- Privacy for Free: Posterior Sampling and Stochastic Gradient Monte Carlo
- The Discrete Gaussian for Differential Privacy
- Tempered Sigmoid Activations for Deep Learning with Differential Privacy
- Practical and Private (Deep) Learning without Sampling or Shuffling
- Fast and Memory Efficient Differentially Private-SGD via JL Projections