Deep Learning With Data Privacy via Residual Perturbation

Summary

The paper proposes residual perturbation, a stochastic differential equation-based method for privacy-preserving deep learning that injects Gaussian noise into each residual mapping of ResNets. The authors prove it guarantees differential privacy and reduces the generalization gap, and they report that it is computationally efficient and outperforms DPSGD in utility while preserving membership privacy.