Research map / Federated learning and privacy / Privacy attacks and secure learning
CAFE: Catastrophic Data Leakage in Vertical Federated Learning
Xiao Jin, Pin-Yu Chen, Chia-Yi Hsu, Chia-Mu Yu, Tianyi Chen
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 CAFE rather than prior gradient-leakage attacks | VFL data recovery · CIFAR-10, MNIST, and Linnaeus 5 | PSNR: CIFAR-10: 31.83 vs 18.94 (best baseline); MNIST: 43.15 vs 13.38; Linnaeus 5: 33.22 vs 12.32 | Improved |
| Add the gradient-matching loss term (α) (ablation) | CAFE · Linnaeus 5 | PSNR: 28.62 -> 33.22 | Improved |
| Add the gradient-matching loss term (α) (ablation) | CAFE · CIFAR-10 and MNIST | PSNR: CIFAR-10: 33.93 -> 31.83; MNIST: 31.93 -> 43.15 | Mixed |
| Use a positive truncated-TV threshold ξ rather than ξ = 0 (ablation) | CAFE · CIFAR-10, Linnaeus 5, and MNIST | PSNR: CIFAR-10: 25.57 -> 31.83; Linnaeus 5: 25.29 -> 33.22; MNIST: 34.51 -> 43.15 | Improved |
| Add the TV-norm regularizer (β) (ablation) | CAFE · CIFAR-10 and Linnaeus 5 | PSNR: CIFAR-10: 18.25 -> 31.83; Linnaeus 5: 23.22 -> 33.22 | Improved |
| Add the TV-norm regularizer (β) (ablation) | CAFE · MNIST | PSNR: 31.98 -> 43.15 | No clear effect |
| Add the internal-representation regularizer (γ) (ablation) | CAFE · CIFAR-10, Linnaeus 5, and MNIST | PSNR: CIFAR-10: 12.51 -> 31.83; Linnaeus 5: 12.37 -> 33.22; MNIST: 6.34 -> 43.15 | Improved |
| Increase batch size K from 10 through 100 (ablation) | CAFE · CIFAR-10, MNIST, and Linnaeus 5 | PSNR: CIFAR-10: 30.83, 35.70, 31.83, 36.87, 38.94; MNIST: 32.60, 39.00, 43.15, 47.05, 47.50; Linnaeus 5: 28.00, 30.53, 33.22, | No clear effect |
| Use the single-loop update protocol instead of nested loops (ablation) | CAFE · CIFAR-10, MNIST, and Linnaeus 5 | iterations to stated PSNR stopping point: Single loop: 7300 (8000), 6600 (8000), 12400 (20000) | Improved |
| Use fake gradients instead of ordinary gradients | CAFE attack · CIFAR-10 | PSNR: 28.68 -> 7.67 | Hurt |
| Use fake gradients as the training update signal | VFL model training · CIFAR-10, Linnaeus 5, and MNIST | training performance | No clear effect |
| Increase the learning rate while attacking during training (ablation) | CAFE · CIFAR-10, MNIST, and Linnaeus 5 | PSNR: CIFAR-10: 31.24 -> 27.62 -> 25.22; MNIST: 31.82 -> 28.42 -> 23.60; Linnaeus 5: 30.74 -> 21.45 -> 20.68 | Hurt |
Closest papers by meaning
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning (AAAI 2024 · similarity 0.87)
- Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning (CVPR 2024 · similarity 0.87)
- Federated Unlearning with Gradient Descent and Conflict Mitigation (AAAI 2025 · similarity 0.85)
- VFLAIR: A Research Library and Benchmark for Vertical Federated Learning (ICLR 2024 · similarity 0.85)
- Towards the Robustness of Differentially Private Federated Learning (AAAI 2024 · similarity 0.85)
- Integer Is Enough: When Vertical Federated Learning Meets Rounding (AAAI 2024 · similarity 0.84)
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.
