Research map / Federated learning and privacy / Privacy attacks and secure learning
Dropout Is NOT All You Need to Prevent Gradient Leakage
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 |
|---|---|---|---|
| Apply dropout during training at rate 0.25 | MLP · MNIST | SSIM under IG: 1.00 -> 0.79 | Hurt |
| Apply dropout during training at rate 0.25 | ViT · CIFAR-10 | SSIM under IG: 0.87 -> 0.01 | Hurt |
| Apply dropout at rate 0.25 | ViT · MNIST | test accuracy: 98.76 -> 98.98 | Helped |
| Apply dropout at rate 0.25 | ViT · CIFAR-10 | test accuracy: 64.47 -> 70.83 | Helped |
| Use the client's dropout masks during the attack | IG · MNIST and CIFAR-10 | SSIM: 0.04 -> 0.99; 0.01 -> 0.93 | Helped |
| Jointly optimize dummy data and dropout masks | IG · ImageNet | SSIM: 0.01 -> 0.72 | Helped |
| Jointly optimize dummy data and dropout masks | IG · CIFAR-10 | SSIM: 0.28 -> 0.71 | Helped |
| Add mask regularization that matches the known dropout rate (ablation) | DIA · MNIST and CIFAR-10 | SSIM | Helped |
| Replace standard mask initialization with analytical mask reconstruction (ablation) | DIA on MLP · CIFAR-10 | SSIM: 0.95 ($\pm$0.04) -> 0.96 ($\pm$0.04) | No clear effect |
| Vary the dropout-mask initialization scheme (ablation) | DIA · CIFAR-10 | SSIM | No clear effect |
| Increase dropout rate and batchsize (ablation) | DIA on MLP and ViT · MNIST and CIFAR-10 | SSIM | Hurt |
| Replace DIA-IG with DIA-APRIL (ablation) | ViT · CIFAR-10 | SSIM: 0.87 ($\pm$0.05) -> 0.82 ($\pm$0.10) | Hurt |
Closest papers by meaning
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning (AAAI 2024 · similarity 0.89)
- Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning (CVPR 2024 · similarity 0.88)
- High-Fidelity Gradient Inversion in Distributed Learning (AAAI 2024 · similarity 0.88)
- Mjölnir: Breaking the Shield of Perturbation-Protected Gradients via Adaptive Diffusion (AAAI 2025 · similarity 0.88)
- Foreseeing Reconstruction Quality of Gradient Inversion: An Optimization Perspective (AAAI 2024 · similarity 0.87)
- Gradient Inversion Attacks on Parameter-Efficient Fine-Tuning (CVPR 2025 · 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.