Research map / Federated learning and privacy / Differential privacy
Spectral Gradient Orthogonalization Improves Differentially Private Training at Scale
Sabari Shanmugam, Nick Barnes, Kerry Taylor
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spectral orthogonalization (DP-Muon) in place of plain DP-SGD update at large batch | DP-SGD on WRN-16-4 · CIFAR-10 (eps=4, B=4096) | test accuracy (%): 39.3 -> 48.7 | Improved |
| add spectral orthogonalization at small batch size (low SNR) | DP-SGD on WRN-16-4 · CIFAR-10 (eps=4, B=256) | test accuracy (%): 39.4 -> 39.0 | No clear effect |
| add spectral orthogonalization at small batch on CIFAR-100 | DP-SGD on WRN-16-4 · CIFAR-100 (eps=4, B<=1024) | test accuracy (%): 26.6 -> 22.0 | Hurt |
| add spectral orthogonalization at very large batch on CIFAR-100 | DP-SGD on WRN-16-4 · CIFAR-100 (eps=4, B=8192) | test accuracy (%): 21.6 -> 25.8 | Improved |
| add spectral orthogonalization on SVHN with large batch | DP-SGD on WRN-16-4 · SVHN (eps=4, B=4096) | test accuracy (%): 20.3 -> 35.2 | Improved |
| add orthogonalization with larger model (WRN-28-10) | DP-SGD on WRN-28-10 · CIFAR-10 (eps=4, B=4096) | test accuracy (%): 34.0 -> 54.9 | Improved |
| add orthogonalization on ResNet-18 | DP-SGD on ResNet-18 · CIFAR-10 (eps=4, B=4096) | test accuracy (%): +14.9% | Improved |
| add orthogonalization to temporal filtering (DiSK-Muon vs DiSK-SGD) | DiSK-SGD · CIFAR-10 (eps=4, B=4096) | test accuracy (%): 39.9 -> 45.8 | Improved |
| add orthogonalization to DOPPLER low-pass filtering | DOPPLER-SGD · CIFAR-10 (eps=4, B=4096) | test accuracy (%): 32.3 -> 50.3 | Improved |
| add temporal Kalman filtering at small batch | DP-SGD · CIFAR-10 (eps=4, B=256) | test accuracy (%): 39.4 -> 49.0 | Improved |
| use magnitude-preserving scaling (DP-Muon-S) instead of pure orthogonalization at high SNR (ablation) | DP-Muon · CIFAR-10 (eps=4, B=4096) | test accuracy (%): 48.7 -> 46.3 | Hurt |
| add orthogonalization in full fine-tuning at large batch | DP-SGD on ViT-Small · CIFAR-100 full fine-tuning (eps=4, B=4096) | test accuracy (%): 79.5 -> 89.3 | Improved |
Closest papers by meaning
- Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization (NeurIPS 2026 · similarity 0.86)
- SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLM Training (NeurIPS 2025 · similarity 0.86)
- Understanding Gradient Orthogonalization for Deep Learning via Non-Euclidean Trust-Region Optimization (NeurIPS 2026 · similarity 0.86)
- Optimal Rates for Vector-Valued Spectral Regularization Learning Algorithms (NeurIPS 2024 · similarity 0.86)
- The Implicit Bias of Heterogeneity towards Invariance: A Study of Multi-Environment Matrix Sensing (NeurIPS 2024 · similarity 0.85)
- Can Entry-Wise Clipping Give Spectral Control of Stochastic Gradients? (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
