Research map / Optimization / Optimizers
Deep Leakage from Gradients
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 DLG to reconstruct training examples from shared gradients | DLG · LFW, Eyeglasses | AUC: 0.94 -> 1.00 | Helped |
| Use DLG to reconstruct training examples from shared gradients | DLG · LFW, Asian | AUC: 0.92 -> 1.00 | Helped |
| Increase communication frequency from 1 to 4 iterations | DLG in federated learning · Federated learning | MSE: $3.3 \times 10^{-6}$ -> $1.8 \times 10^{-2}$ | Hurt |
| Attack gradients at different points during training | DLG · Training stages, 0% to 100% epochs | MSE: $5.7 \times 10^{-6}$; $3.1 \times 10^{-7}$; $4.4 \times 10^{-6}$; $3.3 \times 10^{-6}$ | No clear effect |
| Add Gaussian noise at $10^{-4}$ or $10^{-3}$ | Shared gradients on CIFAR-100 · CIFAR-100 | defendability: Accuracy: 76.3\% -> 75.6\%; 73.3\% | No clear effect |
| Add Gaussian noise at $10^{-2}$ or $10^{-1}$ | Shared gradients on CIFAR-100 · CIFAR-100 | defendability: Accuracy: 45.3\%; $\le$1\% | Helped |
| Add Laplacian noise at $10^{-4}$ or $10^{-3}$ | Shared gradients on CIFAR-100 · CIFAR-100 | defendability: Accuracy: 76.3\% -> 75.6\%; 73.4\% | No clear effect |
| Add Laplacian noise at $10^{-2}$ or $10^{-1}$ | Shared gradients on CIFAR-100 · CIFAR-100 | defendability: Accuracy: 46.2\%; $\le$1\% | Helped |
| Convert shared gradients to FP-16 | Shared gradients on CIFAR-100 · CIFAR-100 | defendability: Accuracy: 76.3\% -> 76.1\% | No clear effect |
| Quantize shared gradients to 8 bit | Shared gradients on CIFAR-100 · CIFAR-100 | defendability: Accuracy: 76.3\% -> 53.7\% | Helped |
| Use L-BFGS rather than SGD for reconstruction | DLG · Nvidia Tesla v100 | reconstruction time: 30 minutes -> 5 minutes | Helped |
| Increase batch size | DLG · Multi-batch examples | reconstruction quality | Mixed |
Closest papers by meaning
- High-Fidelity Gradient Inversion in Distributed Learning (AAAI 2024 · similarity 0.86)
- Revisiting Gradient Pruning: A Dual Realization for Defending against Gradient Attacks (AAAI 2024 · similarity 0.86)
- Differentially private optimization for non-decomposable objective functions (ICLR 2025 · similarity 0.85)
- Certified Machine Unlearning via Noisy Stochastic Gradient Descent (NeurIPS 2024 · similarity 0.85)
- Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning (NeurIPS 2024 · similarity 0.85)
- Mjölnir: Breaking the Shield of Perturbation-Protected Gradients via Adaptive Diffusion (AAAI 2025 · 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.