Research map / Adversarial robustness and security / Attacks on text and graphs
Measuring Data Reconstruction Defenses in Collaborative Inference Systems
Mengda Yang, Ziang Li, Juan Wang, Hongxin Hu, Ao Ren, Xiaoyang Xu, Wenzhe Yi
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 |
|---|---|---|---|
| replace ERA decoder input with SFD-distilled features | Inverse-Network attack · CelebA 2nd conv block, adversarial defense | MSE: 0.124 -> 0.032 | Improved |
| replace ERA with SFD | Inverse-Network attack · CelebA 2nd conv block, adversarial defense | SSIM: 0.370 -> 0.618 | Improved |
| SFD against noise mask defense | SFD attack · CelebA 2nd conv block | MSE: 47.7% | Improved |
| SFD against dropout defense | SFD attack · CelebA 2nd conv block | MSE: 40.9% | Improved |
| move split point from 2nd conv block to avgpool | SFD attack gain over ERA · CelebA, adversarial defense | MSE improvement: 0.092 -> 0.003 | Hurt |
| move split point from 2nd conv block to avgpool | SFD attack gain over ERA · CelebA, adversarial defense | SSIM improvement: 0.248 -> 0.015 | Hurt |
| split at 1st fc layer instead of conv block | SFD attack · CelebA, adversarial defense | SSIM: -0.002 | No clear effect |
| increase adversarial lambda from 1.0 to 5.0 | ERA attack · CIFAR10 1st conv block | SSIM: 0.945 -> 0.309 | Hurt |
| increase adversarial lambda from 1.0 to 5.0 | target VGG-16 classifier · CIFAR10 1st conv block | accuracy: 83.1 -> 76.5 | Hurt |
| increase noise scale b from 1.0 to 10.0 | target VGG-16 classifier · CIFAR10 1st conv block | accuracy: 85.5 -> 82.3 | Hurt |
| increase dropout rate r from 0.5 to 0.9 | target VGG-16 classifier · CIFAR10 1st conv block | accuracy: 84.9 -> 79.9 | Hurt |
| SFD under strongest defense hyper-parameters | SFD attack · CIFAR10 1st conv block | SSIM improvement over ERA: 0.181 | Improved |
Closest papers by meaning
- Reimagining Mutual Information for Enhanced Defense against Data Leakage in Collaborative Inference (NeurIPS 2024 · similarity 0.90)
- A Stealthy Wrongdoer: Feature-Oriented Reconstruction Attack against Split Learning (CVPR 2024 · similarity 0.86)
- Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems (CVPR 2025 · similarity 0.86)
- Unifying Re-Identification, Attribute Inference, and Data Reconstruction Risks in Differential Privacy (NeurIPS 2025 · similarity 0.86)
- Revisiting Gradient Pruning: A Dual Realization for Defending against Gradient Attacks (AAAI 2024 · similarity 0.86)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · 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.
