Scaling up Differentially Private Deep Learning with Fast Per-Example Gradient Clipping

Recent work on Renyi Differential Privacy has shown the feasibility of\napplying differential privacy to deep learning tasks. Despite their promise,\nhowever, differentially private deep networks often lag far behind their\nnon-private counterparts in accuracy, showing the need for more research in\nmodel architectures, optimizers, etc. One of the barriers to this expanded\nresearch is the training time -- often orders of magnitude larger than training\nnon-private networks. The reason for this slowdown is a crucial privacy-related\nstep called "per-example gradient clipping" whose naive implementation undoes\nthe benefits of batch training with GPUs. By analyzing the back-propagation\nequations we derive new methods for per-example gradient clipping that are\ncompatible with auto-differentiation (e.g., in PyTorch and TensorFlow) and\nprovide better GPU utilization. Our implementation in PyTorch showed\nsignificant training speed-ups (by factors of 54x - 94x for training various\nmodels with batch sizes of 128). These techniques work for a variety of\narchitectural choices including convolutional layers, recurrent networks,\nattention, residual blocks, etc.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC