Sliced Rényi Pufferfish Privacy: Directional Additive Noise Mechanism and Private Learning with Gradient Clipping

The researchers propose the Sliced Rényi Pufferfish Privacy (SRPP) privacy framework, replacing high-dimensional Rényi divergence with a projection-based sliced metric to address the curse of dimensionality of RPP and the lack of composition tools for iterative learning. A companion sliced Wasserstein mechanism achieves closed-form Gaussian noise calibration, and the authors propose SRPP-SGD with gradient clipping and privacy accountants based on History-Uniform Caps (HUC) and its subsampled variant sa-HUC. Static and iterative privatization experiments show that the framework has a good privacy-utility trade-off and practical scalability.