Reproduce “Disparate Impact in Differential Privacy from Gradient Misalignment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maria S. Esipova, Atiyeh Ashari Ghomi, Yaqiao Luo, Jesse C. Cresswell
Research from Layer 6 AI pinpoints gradient misalignment from inequitable clipping as the primary cause of disparate impact in differentially private stochastic gradient descent (DPSGD). Their proposed method, DPSGD-Global-Adapt, leverages adaptive global gradient scaling to mitigate this unfairness and improve utility without requiring sensitive group labels, achieving reduced fairness gaps and higher accuracy on various datasets.
Key insights
- Gradient misalignment, specifically direction error caused by inequitable gradient clipping in DPSGD, is the most significant source of disparate impact, disproportionately affecting underrepresented groups.
- Preserving the relative directions of gradients through global scaling methods is more critical for mitigating disparate impact and improving utility than solely bounding individual gradient norms.
- It is possible to significantly reduce disparate impact and improve fairness in differentially private learning without requiring access to protected group labels during training, thereby avoiding disparate treatment and privacy risks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.