AI RESEARCH
Deep Learning under Fractional-Order Differential Privacy
arXiv CS.LG
•
ArXi:2605.09890v1 Announce Type: cross Differentially private stochastic gradient descent (DP-SGD) is a standard approach to privacy-preserving learning based on per-example clipping, subsampling, Gaussian perturbation, and privacy accounting. Classical DP-SGD releases a noisy version of the current clipped subsampled gradient sum.