Revisiting Privacy Amplification by Subsampling in Selective Release DPSGD
Researchers have identified a flaw in the privacy accounting of the Differentially Private Selective Update and Release (DPSUR) algorithm. The existing method overlooks variations in sampling probability introduced by its selective release mechanism, potentially weakening privacy guarantees. To address this, a new algorithm called Differentially Private Selective Release based on Clipped Gradients (DPSR-CG) has been proposed, which offers a more rigorous privacy analysis and demonstrates strong performance across various datasets. AI
IMPACT Enhances privacy guarantees for machine learning models trained on sensitive data, potentially enabling wider adoption in regulated industries.