Gradients Look Alike: Sensitivity is Often Overestimated in DP-SGD

Watch a research presentation from USENIX Security '24 exploring how differentially private stochastic gradient descent (DP-SGD) provides stronger privacy guarantees than previously thought for many datapoints in common benchmark datasets. Learn about a novel per-instance privacy analysis that demonstrates how points with similar neighbors in datasets enjoy better data-dependent privacy protection compared to outliers. Discover how researchers developed a new composition theorem to analyze entire training runs, formally proving that DP-SGD leaks significantly less information than indicated by current data-independent guarantees when training on standard benchmarks. Understand the implications for privacy attacks and how they may fail against many datapoints without sufficient adversarial control over training datasets.