Reproduce “PRISM: Gauge-Invariant Tangent-Space Differentially Private LoRA” 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

Shihao Wang, Xueru Zhang

PRISM (Projected Riemannian Invariant Subspace Mechanism) introduces an intrinsic differentially private mechanism for LoRA fine-tuning that resolves issues of gauge sensitivity and noise amplification found in existing factor-space DP methods. It achieves improved utility across various NLU and reasoning tasks under strong privacy budgets, obtaining an average performance of 0.690 at ε=6 and 0.680 at ε=3 on GLUE8 and Math-10K benchmarks, while demonstrating robustness across model backbones and LoRA ranks.

Key insights

  • Aligning the differential privacy mechanism with the intrinsic geometry and gauge symmetry of LoRA updates (by operating in the tangent space) is crucial for stable and consistent privacy protection.
  • Effective intrinsic noise must be additive, isotropic, and gauge-invariant to avoid arbitrary amplification and spurious higher-order terms that can arise from naive factor-space noising.
  • Adaptive optimizers in DP-LoRA require specific modifications, such as DP-aware floors for preconditioners, to prevent the effective amplification of DP noise and ensure numerical stability when handling low-rank matrices.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.