Research map / Detection and segmentation / Open-set and few-shot detection
Membership Inference Attacks Are Easier on Difficult Problems
Avital Shafran, Shmuel Peleg, Yedid Hoshen
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 |
|---|---|---|---|
| subtract single-sample predictability error from reconstruction error (ablation) | Pix2pix · Facades | MIA ROCAUC (%): 93.62% -> 97.59% | Improved |
| subtract predictability error from reconstruction error (ablation) | Pix2pixHD · Cityscapes | MIA ROCAUC (%): 96.04% -> 99.09% | Improved |
| subtract predictability error from reconstruction error (ablation) | SPADE · ADE20K | MIA ROCAUC (%): 85.31% -> 89.79% | Improved |
| subtract predictability error from reconstruction error (ablation) | Pix2pix · Maps2sat | MIA ROCAUC (%): 84.22% -> 85.65% | Improved |
| subtract predictability error from reconstruction error (ablation) | PraNet · Polyp | MIA ROCAUC (%): 96.03% -> 96.38% | Improved |
| use harder task with more output uncertainty (landmarks to face instead of auto-encoding) | reconstruction-based MIA · CelebA tasks | MIA ROCAUC (%): 50.74% -> 99.54% | Improved |
| increase number of output pixels used by the attack (1 to 200) | reconstruction-based MIA · Pix2pixHD, UperNet, HRNetV2 | MIA ROCAUC (%) | Improved |
| replace human-supervised difficulty score with self-supervised single-sample predictability error | L_mem · Pix2pix Facades | MIA ROCAUC (%): 93.67% -> 97.59% | Improved |
| use in-distribution multi-image predictability error (100 samples) instead of single sample | L_mem · Maps2sat | MIA ROCAUC (%): 85.65% vs 92.43% (Pix2pix); 99.42% vs 82.87% (Pix2pixHD) | Mixed |
| replace single-sample predictability with BDD-trained shadow predictability | L_mem · Cityscapes | MIA ROCAUC (%): 99.29% -> 66.2% (Pix2pixHD); 82.94% -> 74.43% (Pix2pix) | Hurt |
| use shadow-model classifier attack instead of L_mem | MIA · Maps2sat Pix2pix | MIA ROCAUC (%): 85.65% -> 80.15% | Hurt |
| add reconstruction loss to victim training (vs cGAN loss only) (ablation) | reconstruction-based MIA on Pix2pixHD · Cityscapes | MIA ROCAUC (%): 56.58% -> 96.04% | Improved |
Closest papers by meaning
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · similarity 0.88)
- Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models (CVPR 2025 · similarity 0.87)
- Towards Black-Box Membership Inference Attack for Diffusion Models (ICML 2025 · similarity 0.87)
- Causal Evaluation of Membership Inference Attacks (NeurIPS 2026 · similarity 0.87)
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 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.
