🤖 machine learning

Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning

This paper identifies and addresses "localized collateral forgetting," a phenomenon where machine unlearning methods disproportionately degrade predictions for training examples near the deleted data due to inconsistent surrogate targets, proposing "Local Teacher Distillation" as an effective mitigation strategy that aligns unlearned models more closely with full retraining.

Original authors: Polina Dolgova, Sebastian U. Stich

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Polina Dolgova, Sebastian U. Stich

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart student who has studied a massive textbook containing thousands of examples. One day, a law says you must remove a specific chapter from that textbook and make sure the student "forgets" it completely.

The Goal: You want the student to unlearn that chapter without having to re-read the entire book from scratch (which would take too long). You want them to perform just as well on the rest of the book as if they had simply skipped that chapter during their initial study.

The Problem: "Collateral Damage"
The paper discovers a hidden flaw in how we usually try to make students "forget."

When researchers try to force a student to forget a specific topic, they often use a blunt instrument: they either tell the student to "un-study" that topic (pushing the knowledge away) or they tell the student that the topic is nonsense (assigning random, wrong answers).

The paper shows that this approach causes Localized Collateral Forgetting. Think of it like this:

If you tell a student, "Forget everything about Couches," and you do it aggressively, the student might start getting confused about Sofas, Armchairs, and Furniture in general.
Even though those other items were not in the "delete" list, the student's understanding of them gets messed up because they are so similar to the thing they were told to forget.
The student ends up performing worse on these "nearby" topics than they would have if they had just skipped the chapter and kept studying the rest of the book normally.

Why Does This Happen?
The authors explain that the student's brain (the AI model) connects similar ideas together. When you force the student to forget one specific idea using a "random" or "wrong" answer, that confusion spreads to the neighboring ideas in their mental map. It's like tripping over a rug; you don't just fall on the rug, you might knock over the lamp right next to it.

The Solution: The "Local Teacher"
Instead of just telling the student "this is wrong" or "forget this," the authors propose a smarter strategy called Local Teacher Distillation.

Here is the analogy:

The Problem: You can't ask the student to re-read the whole book to see how they should have handled the "Couch" chapter if they had just skipped it. That's too expensive.
The Fix: You hire a tiny, specialized tutor (a "Local Teacher").
How it works: This tutor only looks at the neighbors of the "Couch" chapter—things like "Sofas" and "Armchairs" that the student still remembers. The tutor studies these safe examples and figures out, "Okay, if we didn't have the 'Couch' chapter, how would we logically answer questions about Sofas?"
The Result: The tutor gives the student a "soft" hint (a gentle, correct prediction) on what to do with the "Couch" chapter, based on what they know about the neighbors. This is much better than telling the student to guess randomly.

What They Found
The researchers tested this on a computer system learning to recognize images (like cats, dogs, and couches).

Old Way (Random Guessing): When they tried to delete images of "couches," the system got confused about other furniture nearby.
New Way (Local Teacher): By using the tiny tutor to give gentle hints based on similar, safe images, the system forgot the "couches" just as well as the old way, but it didn't get confused about the other furniture. It stayed much closer to the performance of a student who had simply skipped the chapter from the start.

In Summary
The paper argues that when we try to make AI "forget," we often accidentally hurt its knowledge of similar things. By using a small, smart helper to guide the forgetting process with gentle, logical hints instead of random noise, we can make the AI forget exactly what it needs to, without damaging the rest of its knowledge.

Technical Summary: Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning

Problem Statement
Machine unlearning aims to remove the influence of specific training examples (the "forget set") from a trained model without the computational cost of full retraining. While standard evaluations rely on aggregate metrics (e.g., retain accuracy, forget accuracy, and privacy audit scores), the authors argue these summaries can mask localized failures. Specifically, an unlearned model may appear close to a retrained reference on average while exhibiting sharp deviations on specific retained or test examples.

The paper identifies a failure mode termed localized collateral forgetting: unintended degradation in model predictions on non-deleted examples that are geometrically close to the forget set in representation space. This occurs because examples are coupled through shared representations; interventions designed to suppress the forget-set signal (such as gradient ascent or random labeling) propagate perturbations to nearby retained points.

Methodology
The authors approach the problem through theoretical analysis, empirical observation, and a proposed mitigation strategy.

Theoretical Analysis:
- Gradient Ascent (GA): In linear regression, the authors prove that the excess loss of an unlearned model relative to a retrained model scales with the squared alignment ( $\langle x_q, x_f \rangle^2$ ) between a retained query point $x_q$ and the deleted point $x_f$ .
- Random Labeling (RL): In logistic regression, the discrepancy between RL and retraining is shown to be governed by the projection of the query point onto the span of the forget set. The random labels imposed on the forget set create a "target mismatch" that conflicts with the local prediction structure induced by retraining. This mismatch propagates to nearby points.
- Retain Fine-Tuning: The analysis demonstrates that while fine-tuning on the retain set can restore agreement on retained training points, it does not necessarily correct the model's behavior along components lying within the forget span. Thus, target mismatch remains a root cause.
Proposed Method: Local Teacher Distillation (LTD)
Motivated by the finding that target mismatch drives localized errors, the authors propose Local Teacher Distillation. Instead of assigning random or corrupted labels to the forget set, LTD replaces them with "soft labels" derived from a small auxiliary teacher model.
- Support Selection: The teacher is trained only on a subset of retained examples ( $S_k$ ) that are nearest neighbors to the forget set in representation space (measured by cosine similarity of penultimate-layer embeddings). This ensures the teacher approximates the local prediction structure of the retrained model near the forget set.
- Training Objective: The unlearning process optimizes a loss function that preserves performance on the retain set (using hard labels) while minimizing the cross-entropy between the model's predictions on the forget set and the teacher's soft predictions.
- Efficiency: The teacher is a smaller architecture trained only on the selected support set, avoiding the cost of full retraining.

Key Contributions

Phenomenon Identification: The paper establishes that collateral forgetting is not uniform but concentrated on non-forget examples close to the forget set in representation space.
Theoretical Characterization: The authors provide theoretical proofs for GA and RL methods showing that the deviation from retraining is governed by the alignment or projection of retained points onto the forget set. They identify target mismatch (inconsistent surrogate targets vs. retraining structure) as the underlying mechanism.
Mitigation Strategy: They introduce Local Teacher Distillation, a simple method that uses a local teacher to generate consistent soft labels for the forget set.
Empirical Validation: Experiments on CIFAR-100 (partial-class deletion) show that LTD substantially reduces the discrepancy from retraining, particularly for affected-class and high-similarity examples, while maintaining competitive aggregate unlearning metrics.

Results

Aggregate Metrics: On CIFAR-100 with 50% class deletion, LTD achieves the lowest "Avg. Gap" (mean absolute deviation from retraining across accuracy and privacy metrics) among compared baselines (GA, RL, FT, SalUn, IU, AMUN).
Localized Performance: While aggregate metrics can be misleading, LTD significantly outperforms baselines on affected-class subsets (retained and test points belonging to the deleted class). For instance, while other methods showed large drops in accuracy for these specific points relative to retraining, LTD maintained stability.
Similarity Analysis: The paper demonstrates that competing methods exhibit sharp accuracy drops as similarity to the forget set increases. In contrast, LTD maintains low deviation across all similarity bins.
Ablation Studies: The results confirm that using nearest-neighbor support for the teacher is critical; random or distant support fails to capture the relevant prediction structure, leading to poor alignment with retraining.

Significance and Claims
The paper claims that standard unlearning evaluations are insufficient because they hide structured failures near the forget set. The significance of this work lies in:

Revealing that target mismatch is a fundamental cause of localized collateral forgetting, persisting even with retain-set fine-tuning.
Demonstrating that improving the local consistency of forget-set targets (via a local teacher) is a more effective mitigation strategy than relying solely on retain-set recovery.
Providing a practical, low-cost method (LTD) that brings unlearned models substantially closer to the retraining reference, particularly in the critical neighborhood of the deleted data, without sacrificing privacy or utility.

The authors remain modest regarding generalization, noting that their method relies on the quality of the local support set and representation. They acknowledge that highly heterogeneous forget sets may require clustering or multiple local teachers, and that extension to broader architectures and datasets remains future work.

Technical Summary: Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning

More like this