AI

Data Unlearning via Inverse Distillation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 10:18:25 UTC.

Briefings

Researchers propose Inverse Distillation Unlearning (IDU), a unified framework to remove specific training data from multi-step matching models while retaining model quality. The method formulates distillation as a min-max objective, representing the data distribution as a mixture of the forget-set and generated distributions. By comparing this mixture to the teacher's training distribution, the approach recovers only the retained data at the optimum. Experiments on MNIST and CIFAR-10 under flow-matching and score-based diffusion settings show IDU significantly reduces forgotten class generation frequency while preserving quality on retained classes.

Key points

• IDU is the first unified framework for simultaneous unlearning and distillation in unconditional flow-matching and score-based models.
• The method requires only a pretrained full-data teacher and data from the forget set, without access to retained training examples or extra feature extractors.

Context

This research addresses the trade-off between model efficiency and data privacy in generative AI, offering a new approach to unlearning specific training subsets without compromising the model's overall performance on retained data.

Limitations

The study covers MNIST and CIFAR-10 datasets under flow-matching and score-based diffusion settings. Results are based on extensive experiments but may not generalize to all model architectures or data distributions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/10/2026, 19:53:26 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibpsdqk7eibifx6qwggf63ufm6r6epmrdrfjvxsjy7xi4spvos5ru