Research map / Federated learning and privacy / Privacy attacks and secure learning
Enhancing Privacy Preservation in Federated Learning via Learning Rate Perturbation
Guangnian Wan, Haitao Du, Xuejing Yuan, Jun Yang, Meiling Chen, Jie Xu
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 |
|---|---|---|---|
| perturb local learning rates uniformly at random (LRP) instead of fixed rates (ablation) | FedAvg (ResNet-18) · CIFAR-10 (IID) | test accuracy (%): 94.30 -> 94.15 | No clear effect |
| perturb local learning rates uniformly at random (LRP) (ablation) | FedAvg (ResNet-18) · CIFAR-100 (non-IID) | test accuracy (%): 65.69 -> 65.44 | No clear effect |
| perturb local learning rates uniformly at random (LRP) (ablation) | FedAvg (ResNet-18) · CIFAR-100 (IID) | test accuracy (%): 72.15 -> 72.65 | No clear effect |
| add personalized LR expectations by class count (ada-LRP) on top of LRP (ablation) | LRP · CIFAR-10 (non-IID) | test accuracy (%): 93.03 -> 93.32 | Improved |
| add personalized LR expectations (ada-LRP) on top of LRP (ablation) | LRP · CIFAR-100 (non-IID) | test accuracy (%): 65.44 -> 65.69 | Improved |
| give larger LR to client with better data (eta1=0.005, eta2=0.02) instead of the reverse (ablation) | FedAvg on two non-IID MNIST clients (CNN) · MNIST non-IID 2 clients | test accuracy (%): 85.97 -> 94.21 | Improved |
| use random uniform local LRs with same expectations on two-client MNIST (ablation) | FedAvg (CNN, MLP, logistic regression) · MNIST non-IID 2 clients | test accuracy (%): 92.10 -> 92.00 | No clear effect |
| use LRP defense instead of Soteria, noise, clipping, compression baselines | FL on ImageNet vs GGL attack · ImageNet vs GGL attack | MSE-R: 0.0114 -> 0.3181 | Improved |
| use LRP defense vs Clipping+Noise | FL vs GGL attack · ImageNet vs GGL attack | LPIPS (VGG): 0.6448 -> 0.6966 | Improved |
| add ada-LRP on ImageNet vs vanilla (top-1 vs top-5) (ablation) | FedAvg (RexNet-130) · ImageNet (non-IID) | top-1 accuracy (%): 61.52 -> 61.66; top-5 83.10 -> 83.04 | Mixed |
Closest papers by meaning
- A New Federated Learning Framework Against Gradient Inversion Attacks (AAAI 2025 · similarity 0.91)
- Towards the Robustness of Differentially Private Federated Learning (AAAI 2024 · similarity 0.89)
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning (AAAI 2024 · similarity 0.89)
- Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation (ICML 2025 · similarity 0.89)
- Mjölnir: Breaking the Shield of Perturbation-Protected Gradients via Adaptive Diffusion (AAAI 2025 · similarity 0.88)
- Gradient Inversion Attacks on Parameter-Efficient Fine-Tuning (CVPR 2025 · similarity 0.88)
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.
