Equivariant Differentially Private Deep Learning: Why DP-SGD Needs Sparser Models

Open in new window