Research map / Federated learning and privacy / Differential privacy
Antipodes of Label Differential Privacy: PATE and ALIBI
Mani Malek Esmaeili, Ilya Mironov, Karthik Prasad, Igor Shilov, Florian Tramer
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 ALIBI instead of LP-2ST at eps=3 | label DP training · CIFAR-100 ResNet18 | test accuracy: 28.7 -> 55.0 | Improved |
| Use ALIBI instead of LP-2ST at eps=4 | label DP training · CIFAR-100 ResNet18 | test accuracy: 50.2 -> 65.0 | Improved |
| Use ALIBI instead of LP-2ST at eps=5 | label DP training · CIFAR-100 ResNet18 | test accuracy: 63.5 -> 69.0 | Improved |
| Use ALIBI instead of LP-2ST at eps=8 | label DP training · CIFAR-100 ResNet18 | test accuracy: 74.1 -> 74.4 | No clear effect |
| Replace Laplace noise with Gaussian noise (AGIBI) (ablation) | ALIBI · CIFAR-10 High accuracy | theoretical epsilon: 8.0 -> 19 | Hurt |
| Replace Laplace noise with Gaussian noise (AGIBI) (ablation) | ALIBI · CIFAR-100 High accuracy | theoretical epsilon: 6.3 -> 17 | Hurt |
| Replace Laplace noise with Gaussian noise (AGIBI) (ablation) | ALIBI · CIFAR-100 Medium accuracy | theoretical epsilon: 3.0 -> 8.4 | Hurt |
| Replace Laplace noise with Gaussian noise (AGIBI) (ablation) | ALIBI · CIFAR-10/100 | empirical privacy loss eps_m: 2.9--4.0 -> 2.1--3.2 | No clear effect |
| Insert 1000 canaries in one model vs 10 runs of 100 canaries (ablation) | memorization attack heuristic · CIFAR-100 High, ALIBI | eps_m 95% CI: 2.8--3.5 -> 2.7--3.4 | No clear effect |
| Insert 1000 canaries vs 10x100 canaries (ablation) | memorization attack heuristic · CIFAR-10 Medium, ALIBI | eps_m 95% CI: 1.0--2.2 -> 0.2--2.5 | Mixed |
| Use PATE-FM instead of ALIBI | label DP training · CIFAR-100 High | empirical privacy loss eps_m: 2.8--3.5 -> 0.4--1.0 | Improved |
| Use PATE-FM instead of ALIBI | label DP training · CIFAR-100 High | theoretical epsilon: 6.3 -> 715 | Hurt |
Closest papers by meaning
- Enhancing Learning with Label Differential Privacy by Vector Approximation (ICLR 2025 · similarity 0.87)
- Beyond the Calibration Point: Mechanism Comparison in Differential Privacy (ICML 2024 · similarity 0.86)
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.86)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.85)
- Auditing Privacy Mechanisms via Label Inference Attacks (NeurIPS 2024 · similarity 0.85)
- Provable Privacy with Non-Private Pre-Processing (ICML 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.
