Research map / Label-efficient and robust learning / Recognition and long-tail
Label-Agnostic Forgetting: A Supervision-Free Unlearning in Deep Models
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 Label-Agnostic Forgetting (LAF), a supervision-free machine unlearning method that uses two VAEs over extractor representations plus extractor-unlearning and contrastive representation-alignment losses, with optional supervised repair.
Problem
Machine unlearning methods typically require labels for the forgetting and remaining data during unlearning, which is impractical when annotations are scarce or unavailable at unlearning time.
Approach
Unlearn at the representation level: fit two VAEs to the extractor's representations of remaining and forgetting data, push the extractor's representations of forgetting data away from the forgetting VAE while keeping remaining data fitting the full-data VAE, and use a contrastive loss that aligns post-unlearning representations with the original model's to preserve classifier compatibility withou
Key results
- On data removal with SVHN, LAF reaches Test 92.32% vs Retrain 93.41%, and ASR 57.85% vs Retrain 58.76%, without any labels.
- LAF+R (with supervised repair) achieves the highest performance in 19 of 44 evaluations, versus 13 for SISA, per the paper's own count.
- In class removal, LAF leaves substantial forgetting-class information: on DIGITS Test_f is 0.26% (close to Retrain's 0%), but its ASR is 52.25% vs 26.49% for Retrain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add supervised repair using labels from remaining data | LAF · data removal, class removal, and noisy label removal | prediction accuracy | Helped |
| add representation alignment loss L_RA (ablation) | LAF · data removal, class removal, and noisy label removal | remaining-data and test accuracy | Helped |
| add extractor unlearning loss L_UE (ablation) | LAF · class removal on DIGITS, FASHION, and CIFAR10 | Test_f accuracy | Helped |
| add extractor unlearning loss L_UE (ablation) | LAF · class removal on SVHN | Test_f accuracy | Mixed |
| include two KL divergence terms in the extractor unlearning objective (ablation) | LAF · data removal | remaining-data and test accuracy | Hurt |
| train VAE h on remaining data D_r instead of full data D (ablation) | LAF · random data removal | prediction performance | No clear effect |
| use two-stage rather than alternating optimization (ablation) | LAF · data removal and class removal | prediction performance | Mixed |
Surprising findings
- Adding KL terms to the extractor unlearning objective greatly hurts data-removal performance, so the paper drops them from the final loss.
- The label-agnostic method underperforms in class removal, failing to fully remove the forgetting class information, which the authors attribute to the lack of labels.
Limitations
- Weaker class removal: the forgetting class is not sufficiently removed without labels.
- Low efficiency: VAE training takes nearly half of the runtime, and LAF is not faster than most baselines.
- Evaluated only on small CNN and ResNet-18 models with four image datasets.
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10 SVHN
Benchmarks: MNIST Fashion-MNIST CIFAR-10 SVHN
Metrics: Train_r accuracy Train_f accuracy Test accuracy membership inference attack success rate
Baselines: Retrain NegGrad Boundary SISA Unrolling T-S SCRUB
Builds on: SCRUB SISA Boundary unlearning Unrolling SGD
Closest papers by meaning
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.90)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.90)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.89)
- CoUn: Empowering Machine Unlearning via Contrastive Learning (NeurIPS 2025 · similarity 0.89)
- Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models (NeurIPS 2024 · similarity 0.88)
- Towards Source-Free Machine Unlearning (CVPR 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.