Research map / Federated learning and privacy / Privacy attacks and secure learning
When Does Data Augmentation Help With Membership Inference Attacks?
Yigitcan Kaya, Tudor Dumitras
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 |
|---|---|---|---|
| Apply augmentation (SL, LS, DL, RC, CO, GA, MU) tuned for maximum accuracy | VGG CNNs · CIFAR-100 | Advpow: varies by mechanism | Mixed |
| Label smoothing for accuracy | VGG CNN · CIFAR-100 | Advpow: Acc +10%, Advpow 60.2 to 75.6 | Hurt |
| Random cropping at high intensity (RAD<25%) (ablation) | VGG CNN · CIFAR-100 | Advpow: 60.2 to 2.9 at P=25 with Acc 42.6% | Improved |
| Random cropping at moderate intensity (RAD<10%) (ablation) | VGG CNN · CIFAR-100 | Advpow: 60.2 to 7.1 at P=14 with Acc 52.8% | Improved |
| Gaussian augmentation | VGG CNN · CIFAR-100 | Advpow: 58.2/60.2 baseline vs 58.8/62.5 at accuracy-tuned | No clear effect |
| Add label smoothing on top of random cropping (ablation) | VGG CNN · CIFAR-100 | Advpow: RC alone P=4 Advpow 32.0 vs 38.4 with α=0.90 | Hurt |
| Label smoothing on non-image tasks (ablation) | fully connected networks · Texas-100 | Adv: 55.0% to 64.6% at α=0.10 | Hurt |
| Label smoothing on non-image tasks (ablation) | fully connected networks · Purchase-100 | Adv: 14.6% to 20.5% at α=0.10 with Acc drop | Mixed |
| Differential privacy (DP-SGD) with eps<=1 | VGG CNN · CIFAR-100 | Acc: 54.3% baseline to 6.8% | Hurt |
| Fewer training epochs (3-7 vs 35) (ablation) | VGG CNN · CIFAR-10 | Adv: 65-90% less Adv, with 5-17% lower Acc | Improved |
Closest papers by meaning
- How Does Data Augmentation Affect Privacy in Machine Learning? (AAAI 2021 · similarity 0.94)
- Causal Evaluation of Membership Inference Attacks (NeurIPS 2026 · similarity 0.88)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.88)
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 2025 · similarity 0.88)
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Privacy Attacks on Image AutoRegressive Models (ICML 2025 · similarity 0.86)
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.
