Geometric Data Perturbation with Noisy-Anchor Alignment for Privacy-Preserving Collaborative Learning
This research addresses privacy risks in collaborative learning using Geometric Data Perturbation (GDP). It proposes Anchor-Aligned Independent GDP with Anchor Noise, which perturbs shared anchor matrices instead of private data representations. This method achieves better accuracy than noise-only methods while maintaining stronger privacy guarantees against analyst collusion by using noisy anchors for alignment.