DP-SGD

DP-SGD injects calibrated noise into gradient updates to guarantee individual data privacy during deep learning training.

Differentially Private Stochastic Gradient Descent (DP-SGD) is the gold standard for training neural networks on sensitive datasets like medical records or private text. By clipping per-sample gradients and adding Gaussian noise (typically controlled by a privacy budget epsilon), it ensures a model cannot leak specific training examples. This technique, popularized by Abadi et al. in 2016, is now integrated into core libraries like PyTorch Opacus and TensorFlow Privacy to protect user data in large-scale AI deployments.

https://arxiv.org/abs/1607.00133