Research map / Federated learning and privacy / Privacy attacks and secure learning
Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data
Ahmed Mehdi Inane, Vincent Quirion, Gintare Karolina Dziugaite, Ioannis Mitliagkas
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 |
|---|---|---|---|
| add 50000 Amazon Reviews public points | Langevin Unlearning · IMDB (LSTM, K=15) | unlearned test accuracy: 82.59 -> 81.42 | Hurt |
| flip 40% of public labels (more mismatch) versus clean public data (ablation) | ALU · IMDB (LSTM, K=15) | unlearned test accuracy: 1.17% -> 2.19% | Hurt |
| flip 40% of public labels versus clean public data (ablation) | ALU retrained model · IMDB (LSTM, K=15) | retrained test accuracy: 0.39% -> 1.74% | Hurt |
| increase public points from 10000 to 40000 with aligned domains (ablation) | ALU · DomainNet aligned (Quickdraw public, Clipart private, K=5) | relative loss difference (%): 4.23 -> 3.68 | No clear effect |
| increase public points from 10000 to 40000 with misaligned domains (ablation) | ALU · DomainNet misaligned (Infograph public, Real private, K=5) | relative loss difference (%): 10.53 -> 10.81 | No clear effect |
| misaligned instead of aligned public-private domains (ablation) | ALU · DomainNet (K=5) | relative loss difference (%): 4.23 -> 10.53 | Hurt |
| add public data to Langevin unlearning | LU · DomainNet Clipart | Renyi divergence to retraining | Improved |
| add public data to Langevin unlearning | LU · IMDB with U-LiRA | U-LiRA confidence | Improved |
Closest papers by meaning
- The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning (ICLR 2025 · similarity 0.93)
- Leveraging Per-Instance Privacy for Machine Unlearning (ICML 2025 · similarity 0.93)
- Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning (NeurIPS 2024 · similarity 0.93)
- A Certified Unlearning Approach without Access to Source Data (ICML 2025 · similarity 0.92)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.91)
- Efficient Source-free Unlearning via Energy-Guided Data Synthesis and Discrimination-Aware Multitask Optimization (ICML 2025 · similarity 0.91)
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.
