Navigating the Challenges of Selective Classification Under Differential Privacy: An Empirical Study

This post discusses the challenges of selective classification under differential privacy and introduces the Selective Classification via Training Dynamics Ensembles (SCTD) method as a novel solution. The method leverages intermediate model checkpoints to mitigate privacy leakage while maintaining competitive performance. It also presents a new evaluation metric that allows for fair comparison of selective classification methods across different privacy levels.