Research map / Label-efficient and robust learning / Recognition and long-tail
Partially Blinded Unlearning: Class Unlearning for Deep Networks from Bayesian Perspective
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Partially-Blinded Unlearning (PBU), a single-step Bayesian class-unlearning loss that uses only the forget-class data plus Fisher-Mahalanobis and l2 parameter-stability regularizers, evaluated on ResNet/All-CNN classifiers across MNIST, CIFAR-10, CIFAR-100 and Food-101.
Problem
Class unlearning in pre-trained classification networks must remove information about a target class while preserving accuracy on the remaining classes, typically under the assumption of full dataset access, which is often unavailable.
Approach
Frames class unlearning as a Bayesian MAP problem, bounds the retain-data log-posterior via a second-order Taylor expansion and Fisher information, and yields a loss that minimizes forget-class log-likelihood with Fisher-Mahalanobis and l2 stability terms, so only forget-class data is needed.
Key results
- PBU reduces forget-class accuracy to near 0 (e.g., 0.1% on MNIST Class-8 with ResNet-18 and 0% on several settings) while retaining higher retain-class accuracy than baselines such as Fast-Effective and Bad Teaching.
- On CIFAR-10/100 the paper reports retain-class accuracy gains of roughly 5% to 20% over baselines, at the cost of marginally higher forget-class accuracy in some settings.
- MIA accuracy of PBU-unlearned models stays below 0.5 across the reported model/dataset settings (e.g., 0.24 on MNIST ResNet-18).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the stability regularizer to PBU (ablation) | PBU · MNIST, CIFAR-10, CIFAR-100, and Food-101 | retain-class accuracy: 20% to 50% | Helped |
| Increase α while holding β and γ at their optimal settings (ablation) | PBU · MNIST, CIFAR-10, CIFAR-100, and Food-101 | forget-class accuracy | Helped |
| Increase α while holding β and γ at their optimal settings (ablation) | PBU · MNIST, CIFAR-10, CIFAR-100, and Food-101 | retain-class accuracy | Hurt |
| Use PBU instead of baseline methods | Class unlearning · CIFAR-10 and CIFAR-100 | retain-class accuracy: 5% to 20% | Helped |
| Use PBU instead of baseline approaches | Class unlearning · Food-101 | retain-class accuracy: 5% to 10% | Helped |
| Use PBU instead of baseline methods | Class unlearning · CIFAR-10 and CIFAR-100 | forget-class accuracy | Hurt |
Surprising findings
- Removing the stability regularizer drives forget accuracy to 0 but collapses retain accuracy by 20% to 50%, showing the regularizer is essential.
Limitations
- Requires only forget-class data but paper gives no hardware or compute details.
- Uses an upper-bound surrogate of the Bayesian objective rather than the intractable lower bound, with no theoretical guarantee of tightness.
- On CIFAR-10 and CIFAR-100 the forget-class accuracy is slightly higher than some baselines (trade-off).
Experimental setup
Datasets: MNIST CIFAR-10 CIFAR-100 Food-101
Benchmarks: MNIST CIFAR-10 CIFAR-100 Food-101
Metrics: forget-class accuracy retain-class accuracy membership inference attack accuracy unlearning time in epochs
Baselines: Retraining Fine-tuning Fast-Effective Bad Teaching
Builds on: elastic weight consolidation Fast Yet Effective unlearning Bad Teaching unlearning
Closest papers by meaning
- Learning to Unlearn: Instance-Wise Unlearning for Pre-trained Classifiers (AAAI 2024 · similarity 0.89)
- Towards Source-Free Machine Unlearning (CVPR 2025 · similarity 0.89)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.89)
- LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty (CVPR 2025 · similarity 0.89)
- Towards Scalable Exact Machine Unlearning Using Parameter-Efficient Fine-Tuning (ICLR 2025 · similarity 0.88)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.88)
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.