Research map / Label-efficient and robust learning / Recognition and long-tail
Efficient Source-free Unlearning via Energy-Guided Data Synthesis and Discrimination-Aware Multitask Optimization
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 DSDA, a two-stage source-free class unlearning framework that synthesizes data via accelerated energy-guided Langevin sampling and unlearns with a discriminative feature alignment loss plus a gradient-conflict-resolving multitask optimizer.
Problem
Machine unlearning of forget classes when the original training data is inaccessible (source-free unlearning), where prior methods rely on costly knowledge distillation and retraining.
Approach
Derive an energy function from classifier logits and synthesize class-conditional data via Langevin dynamics accelerated by Heun (second-order Runge-Kutta) integration and Nesterov momentum; then fine-tune the original model with forget, retain, and a discriminative feature alignment loss, balancing their gradients via a min-norm-style multitask update.
Key results
- DSDA reaches Af of 0% on all single-class tasks and Ar of 77.91% on CIFAR-10 vs 77.90% for Retrain, with ET of 133.40 s vs 1108.10 s for Retrain.
- On CIFAR-100 multi-class unlearning, DSDA keeps Ar around 49.5-50.3% across k=2,4,8 while GKT and LAU degrade sharply as k grows.
- Reports an average 68.50% execution-time improvement over source-free baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AEGDS acceleration instead of unaccelerated EGDS sampling (ablation) | DSDA · CIFAR-10 and CIFAR-100 | Ar and ET | Helped |
| Add the discriminative feature alignment objective (ablation) | DSDA · CIFAR-10 and CIFAR-100 | Ar | Helped |
| Use multitask optimization rather than direct updates with gT (ablation) | DSDA · CIFAR-10 and CIFAR-100 | Ar and ET | Helped |
Surprising findings
- Retrain on CIFAR-10 yields Af 0.00 but DSDA's unlearned model matches its retain accuracy closely, while baselines like ADV+IMP and UNSIR show MIA values above Retrain, indicating forget-data leakage.
Limitations
- Limited to class-level unlearning; not evaluated for instance-wise or attribute-wise unlearning.
- Comparisons with source-free methods are noted as not fully fair, and results are on small models and datasets (ResNet-18, ViT).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 PinsFaceRecognition
Benchmarks: CIFAR-10 CIFAR-100 PinsFaceRecognition
Metrics: Accuracy on forget data (Af) Accuracy on retain data (Ar) membership inference attack (MIA) execution time (ET)
Baselines: Retrain SSD UNSIR ADV+IMP LAU SCAR GKT ISPF
Builds on: SGLD GKT ISPF UNSIR Nesterov momentum
Closest papers by meaning
- The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning (ICLR 2025 · similarity 0.91)
- Towards Source-Free Machine Unlearning (CVPR 2025 · similarity 0.91)
- Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning (NeurIPS 2024 · similarity 0.91)
- Leveraging Per-Instance Privacy for Machine Unlearning (ICML 2025 · similarity 0.91)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.90)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.90)
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.