IMU Lets AI Surgically Forget Private Data Without Touching the Training Set
A retain-data-free machine unlearning method uses influence functions to prioritize which forget samples deserve the strongest updates, boosting utility 30% over gradient ascent.
Takeaways
- IMU is a retain-data-free machine unlearning method that only needs the forget set, not the original training data.
- It uses influence functions to rank which forget samples deserve the strongest unlearning updates.
- A classifier-level Hessian approximation avoids full-model inversion while preserving 0.992 Spearman rank correlation with exact influence.
- On CIFAR-10 class-wise unlearning, IMU keeps 93.38% test accuracy versus 72.43% for gradient ascent at the same forget depth.
- Improves model utility by an average of 30% over uniform gradient ascent across vision and language tasks.
- Head-only updates may leave backbone representations vulnerable to adversarial probing; no public code repository linked yet.
Machine unlearning is the task of surgically removing specific data points from a trained model, ideally as if those points had never been seen. The standard recipe assumes you still have the rest of the training set on hand to keep the model sharp during that surgery. In practice, that assumption often breaks: the retain data may be locked behind privacy rules, discarded for storage reasons, or simply owned by someone else. A new paper from researchers at Harbin Institute of Technology, Monash University, Beihang University, and City University of Hong Kong proposes Influence-guided Machine Unlearning (IMU), a method that works using only the forget set and outperforms existing retain-data-free baselines.
The core insight is that samples in the forget set carry wildly different weight in the trained model. A small subset dominates the model's fit on the forget set, while others contribute marginally, redundantly, or introduce unnecessary drift to utility-relevant parameters when treated uniformly during unlearning. Handling them all identically, as vanilla gradient ascent does, tends to trash the model's overall accuracy.

Why retain-data-free unlearning is hard
Most modern machine unlearning methods fine-tune on the retain set to preserve utility while pushing the forget set out of the model. That works when you have the retain data. Without it, you fall back to methods like gradient ascent (GA), which flips the sign on the loss for the forget samples and climbs the loss surface. GA is brutal: it makes the model forget, but it also degrades everything else because every sample gets the same weight in the update.
Recent retain-data-free methods try to be cleverer. Cha et al. generate adversarial examples with respect to the forgetting data; Bonato et al. use Mahalanobis distance to shift forgetting representations toward incorrect classes followed by knowledge distillation with auxiliary images; and in the LLM domain Negative Preference Optimization (NPO) treats forgetting data as negative examples within a preference optimization framework. Most of these approaches still apply a uniform unlearning objective across the entire forget set. NPO implicitly introduces instance-level weighting through preference optimization, though it scales updates based on output-level prediction confidence rather than a direct measure of a sample's parametric contribution.