Research map / Federated learning and privacy / Differential privacy
PRISM: Gauge-Invariant Tangent-Space Differentially Private LoRA
Shihao Wang, Xueru Zhang
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 |
|---|---|---|---|
| use PRISM tangent-space DP instead of LoRA+ | DP LoRA fine-tuning · Gemma-3-4B-pt 12 tasks, epsilon=6 | average accuracy: 0.674 -> 0.690 | Improved |
| use PRISM instead of LoRA+ | DP LoRA fine-tuning · Gemma-3-4B-pt 12 tasks, epsilon=3 | average accuracy: 0.665 -> 0.680 | Improved |
| use PRISM instead of LoRA-RITE and AdamW without DP | LoRA fine-tuning · Gemma-3-4B-pt 12 tasks, non-DP | average accuracy: 0.782 -> 0.737 (vs LoRA-RITE) | Hurt |
| use PRISM instead of AdamW | DP LoRA · Gemma-2-9B Math-10K, epsilon=6 | average accuracy: 0.6779 -> 0.7019 | Improved |
| use PRISM instead of LoRA+ | DP LoRA · Gemma-3-12B-pt Math-10K, epsilon=6 | average accuracy: 0.6760 -> 0.6966 | Improved |
| use PRISM at LoRA rank 8 (ablation) | DP LoRA · Gemma-3-4B-pt, epsilon=6 | average accuracy: 0.650 -> 0.684 (vs LoRA+) | Improved |
| use PRISM at LoRA rank 32 (ablation) | DP LoRA · Gemma-3-4B-pt, epsilon=6 | average accuracy: 0.653 -> 0.682 (vs LoRA+) | Improved |
| increase rank 16 to 32 in PRISM (ablation) | PRISM · Gemma-3-4B-pt, epsilon=6 | average accuracy: 0.690 -> 0.682 | No clear effect |
| use PRISM tangent-space DP | DP LoRA · step time | step time (s): 9.37 -> 18.64 | Hurt |
| use PRISM tangent-space DP | DP LoRA · peak memory | peak memory (MB): 20961.1 -> 20964.3 | No clear effect |
| use PRISM instead of AdamW on SST-2, RTE, AQuA | DP LoRA · Gemma-3-4B-pt, epsilon=6 | per-task accuracy | Mixed |
| gauge-rescale initialization with PRISM | PRISM · Math-10K gauge sensitivity | range of update norm | Improved |
Closest papers by meaning
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.89)
- Sampling-Free Privacy Accounting for Matrix Mechanisms under Random Allocation (NeurIPS 2026 · similarity 0.88)
- Privacy-Preserving Low-Rank Adaptation Against Membership Inference Attacks for Latent Diffusion Models (AAAI 2025 · similarity 0.87)
- The Privacy Power of Correlated Noise in Decentralized Learning (ICML 2024 · similarity 0.87)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.87)
- Privacy Amplification for Matrix Mechanisms (ICLR 2024 · similarity 0.87)
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.
