Research map / Federated learning and privacy / Differential privacy
Photonic Differential Privacy with Direct Feedback Alignment
Ruben Ohana, Hamlet Jesse Medina Ruiz, Julien Launay, Alessandro Cappelli, Iacopo Poli, Liva Ralaivola, Alain Rakotomamonjy
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 DP noise sigma 0.05 vs 0, clipping kept | DFA · FashionMNIST | test accuracy: 84.20 -> 83.70 | Hurt |
| increase noise sigma 0 to 0.2 | DFA · FashionMNIST | test accuracy: 84.20 -> 81.66 | Hurt |
| increase noise sigma 0 to 0.2 | BP · FashionMNIST | test accuracy: 75.22 -> 66.78 | Hurt |
| activation clipping (non-private to sigma 0) (ablation) | DFA · FashionMNIST | test accuracy: 86.80 -> 84.20 | Hurt |
| activation clipping (non-private to sigma 0) (ablation) | BP · FashionMNIST | test accuracy: 88.33 -> 75.22 | Hurt |
| photonic OPU projection instead of GPU ternarized DFA | PDFA · FashionMNIST | test accuracy: 84.38 -> 83.79 | No clear effect |
| photonic OPU instead of GPU ternarized, sigma 0.2 | PDFA · FashionMNIST | test accuracy: 80.80 -> 80.94 | No clear effect |
| ternarize the error vs plain DFA | TDFA · MNIST | test accuracy: 96,36 -> 97,09 | Improved |
| increase noise sigma 0.01 to 0.1 | BP · MNIST | test accuracy: 62,63 -> 48,33 | Hurt |
| increase noise sigma 0.01 to 0.1 | PDFA · MNIST | test accuracy: 93,60 -> 93,12 | No clear effect |
| add DP sigma 0.05 | BP · CIFAR-10 | test accuracy: 83.17 -> 60.45 | Hurt |
| add DP sigma 0.05 | DFA · CIFAR-10 | test accuracy: 81.34 -> 79.68 | Hurt |
Closest papers by meaning
- Delving into Differentially Private Transformer (ICML 2024 · similarity 0.84)
- DOPPLER: Differentially Private Optimizers with Low-pass Filter for Privacy Noise Reduction (NeurIPS 2024 · similarity 0.83)
- A New Linear Scaling Rule for Private Adaptive Hyperparameter Optimization (ICML 2024 · similarity 0.82)
- Towards hyperparameter-free optimization with differential privacy (ICLR 2025 · similarity 0.82)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.82)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.81)
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.
