Machine Unlearning in Forgettability Sequence

Junjie Chen, Qian Chen, Jian Lou, Xiaoyu Zhang, Kai Wu, Zilong WangarXiv 2024 · OpenReview

TL;DR

This paper tackles the challenge of "machine unlearning" – effectively removing specific data's influence from a trained model. They discover that data points with higher privacy risks are harder to forget and propose a novel framework, RSU, that prioritizes forgetting these difficult-to-remove samples first, significantly improving unlearning performance.

What Problem Does This Paper Address?

The "right to be forgotten" is a growing concern in the age of big data and AI. Regulations like GDPR empower individuals to request the deletion of their personal data. Machine unlearning (MU) aims to achieve this by removing the "trace" of specific data points from a trained model without retraining from scratch (which is computationally expensive). However, a key question remained: are all data points equally "forgettable"? This paper investigates whether different data points have varying levels of difficulty in being unlearned and, crucially, whether the order in which these points are forgotten impacts the overall effectiveness of unlearning algorithms. Understanding this sequence is vital for developing more efficient and robust unlearning methods.

Methods

The core of this paper's contribution lies in its proposed RSU (Ranking and Sequential Unlearning) framework. RSU consists of two main modules:

  1. Ranking Module: This module assesses the "difficulty" of forgetting each data point. The authors found that samples with higher privacy risks (specifically, those more susceptible to model inversion and membership inference attacks) are harder to forget. They use the model's loss value on a data point as a proxy for its privacy risk and thus its unlearning difficulty. The module then ranks the data points from "hardest to forget" to "easiest to forget."
  2. SeqUnlearn Module: This module then processes the unlearning dataset sequentially, following the order determined by the Ranking module. It applies unlearning algorithms to groups of data points, starting with the most difficult ones. This approach mimics human learning strategies, where complex tasks are tackled first. The module also introduces a novel learning objective with a regularizer that dynamically adjusts sample weights based on their difficulty and the learning pace.

Datasets

The authors evaluated their framework on four widely used image classification datasets:

  • MNIST: A dataset of handwritten digits.
  • CIFAR-10: A dataset of 10 common object classes.
  • CIFAR-100: A dataset of 100 common object classes.
  • Tiny-ImageNet: A larger dataset with more diverse images and classes.

They also used three common deep neural network architectures: ResNet-18, ResNet-50, and ResNeXt-50.

Key Insights and Ideas

  • Privacy Risk Correlates with Unlearning Difficulty: The most significant insight is the strong link between a data point's privacy risk (how easily it can be exploited by attacks like model inversion or membership inference) and how difficult it is to remove its influence from a model. This suggests that "easy-to-attack" samples are precisely the ones that are "hard-to-forget."
  • Order Matters Dramatically: The sequence in which data points are unlearned has a substantial impact on the overall effectiveness of the unlearning process. Unlearning in a "hard-to-easy" sequence, as proposed by RSU, consistently outperforms random or "easy-to-hard" sequences.
  • Loss as a Proxy for Difficulty: While directly measuring privacy risk can be complex, the paper demonstrates that the model's loss value on a data point serves as a reliable and practical proxy for its unlearning difficulty. This makes the RSU framework more implementable.
  • "Easy-to-Attack" Samples are Informative Outliers: The paper suggests that samples vulnerable to privacy attacks are often outliers, ambiguous, or underrepresented in the dataset. These characteristics make them more memorable to the model and thus harder to forget.

Key Results

  • RSU Improves Performance Across Methods: The RSU framework consistently improved the performance (measured by accuracy on remaining data and reduced MIA gap) of various baseline unlearning algorithms (GA, RL, Bad-T) across all tested datasets.
  • Significant Accuracy Gains: For example, in class forgetting scenarios, RSU improved the accuracy of the GA algorithm by up to 10% on Tiny-ImageNet compared to random ordering.
  • Reduced MIA Gap: The RSU framework also significantly reduced the Membership Inference Attack (MIA) gap, indicating a more thorough removal of the unlearned data's influence.
  • RSU Outperforms Other Orderings: Ablation studies showed that RSU's "hard-to-easy" ordering significantly outperformed both random ordering and a "reverse" (easy-to-hard) ordering, highlighting the importance of the proposed sequence.

Limitations and Future Work

  • Proxy Limitations: While loss value is a good proxy for unlearning difficulty, it's not a perfect measure of privacy risk. The paper acknowledges that directly quantifying privacy risk for every data point can be computationally intensive.
  • Generalizability of "Easy-to-Attack" Definition: The paper focuses on model inversion and membership inference attacks. Future work could explore other privacy attack vectors and their correlation with unlearning difficulty.
  • Heterogeneous Forget Sets: The paper's insights are particularly strong when dealing with data that has varying levels of unlearning difficulty. Real-world forget sets might be more complex and mixed. Future work could focus on developing strategies for handling such heterogeneous forget sets more effectively.
  • RSU as a General Framework: The authors emphasize that RSU is compatible with most existing unlearning methods, suggesting that future research could integrate RSU with even more advanced unlearning techniques.

Summary updated

Abstract

Machine unlearning (MU) is becoming a promising paradigm to achieve the "right to be forgotten", where the training trace of any chosen data points could be eliminated, while maintaining the model utility on general testing samples after unlearning. With the advancement of forgetting research, many fundamental open questions remain unanswered: do different samples exhibit varying levels of difficulty in being forgotten? Further, does the sequence in which samples are forgotten, determined by their respective difficulty levels, influence the performance of forgetting algorithms? In this paper, we identify key factor affecting unlearning difficulty and the performance of unlearning algorithms. We find that samples with higher privacy risks are more likely to be unlearning, indicating that the unlearning difficulty varies among different samples which motives a more precise unlearning mode. Built upon this insight, we propose a general unlearning framework, dubbed RSU, which consists of Ranking module and SeqUnlearn module.
BibTeX
@misc{chen2024machine,
  author = {Junjie Chen and Qian Chen and Jian Lou and Xiaoyu Zhang and Kai Wu and Zilong Wang},
  title = {{Machine Unlearning in Forgettability Sequence}},
  year = {2024},
  eprint = {2410.06446},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.06446},
}
Figure