Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learn to Forget: Machine Unlearning via Neuron Masking

Yang Liu, Zhuo Ma, Ximeng Liu, Jian Liu, Zhongyuan Jiang, Jianfeng Ma, Philip Yu, Kui Ren

TL;DR

This paper tackles the right-to-forget challenge in neural networks by introducing Forsaken, a neuron-masking unlearning method that uses a mask gradient generator to erase memorization of chosen data. It also defines forgetting rate (FR), a uniform metric based on a membership oracle, to quantify unlearning effectiveness. Forsaken demonstrates strong forgetting performance (average FR > 90%) with minimal utility loss (<5% accuracy drop) across eight diverse datasets, and offers practical efficiency advantages over retraining-based approaches. The approach is shown to generalize beyond neural networks to LR and SVM, and is adaptable to federated learning contexts, enabling scalable, data-subset forgetting with reduced privacy risks from unintended memorization.

Abstract

Nowadays, machine learning models, especially neural networks, become prevalent in many real-world applications.These models are trained based on a one-way trip from user data: as long as users contribute their data, there is no way to withdraw; and it is well-known that a neural network memorizes its training data. This contradicts the "right to be forgotten" clause of GDPR, potentially leading to law violations. To this end, machine unlearning becomes a popular research topic, which allows users to eliminate memorization of their private data from a trained machine learning model.In this paper, we propose the first uniform metric called for-getting rate to measure the effectiveness of a machine unlearning method. It is based on the concept of membership inference and describes the transformation rate of the eliminated data from "memorized" to "unknown" after conducting unlearning. We also propose a novel unlearning method calledForsaken. It is superior to previous work in either utility or efficiency (when achieving the same forgetting rate). We benchmark Forsaken with eight standard datasets to evaluate its performance. The experimental results show that it can achieve more than 90\% forgetting rate on average and only causeless than 5\% accuracy loss.

Learn to Forget: Machine Unlearning via Neuron Masking

TL;DR

This paper tackles the right-to-forget challenge in neural networks by introducing Forsaken, a neuron-masking unlearning method that uses a mask gradient generator to erase memorization of chosen data. It also defines forgetting rate (FR), a uniform metric based on a membership oracle, to quantify unlearning effectiveness. Forsaken demonstrates strong forgetting performance (average FR > 90%) with minimal utility loss (<5% accuracy drop) across eight diverse datasets, and offers practical efficiency advantages over retraining-based approaches. The approach is shown to generalize beyond neural networks to LR and SVM, and is adaptable to federated learning contexts, enabling scalable, data-subset forgetting with reduced privacy risks from unintended memorization.

Abstract

Nowadays, machine learning models, especially neural networks, become prevalent in many real-world applications.These models are trained based on a one-way trip from user data: as long as users contribute their data, there is no way to withdraw; and it is well-known that a neural network memorizes its training data. This contradicts the "right to be forgotten" clause of GDPR, potentially leading to law violations. To this end, machine unlearning becomes a popular research topic, which allows users to eliminate memorization of their private data from a trained machine learning model.In this paper, we propose the first uniform metric called for-getting rate to measure the effectiveness of a machine unlearning method. It is based on the concept of membership inference and describes the transformation rate of the eliminated data from "memorized" to "unknown" after conducting unlearning. We also propose a novel unlearning method calledForsaken. It is superior to previous work in either utility or efficiency (when achieving the same forgetting rate). We benchmark Forsaken with eight standard datasets to evaluate its performance. The experimental results show that it can achieve more than 90\% forgetting rate on average and only causeless than 5\% accuracy loss.

Paper Structure

This paper contains 30 sections, 10 equations, 16 figures, 16 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Machine unlearning
  4. Membership oracle
  5. Forgetting Rate
  6. Forsaken
  7. Intuition
  8. Forsaken in detail
  9. Loss function of $\mathcal{G}$
  10. Non-member data
  11. How to apply Forsaken in applications
  12. Evaluation
  13. Experimental Setup
  14. Performance Comparison
  15. Effect Factors
  16. ...and 15 more sections

Figures (16)

  • Figure 1: The high-level workflow of mask gradient generator.
  • Figure 2: Performance change of the target model over the testing set for OOD data unlearning.
  • Figure 3: Performance change of the target model over the testing set for ID data unlearning.
  • Figure 4: The performance change with different unlearning sizes for OOD data unlearning. Test.Acc means the testing accuracy.
  • Figure 5: The detailed performance evaluation of the mask gradient generator $\mathcal{G}$ at each iteration.
  • ...and 11 more figures

Theorems & Definitions (5)

  • Definition 1: Machine Unlearning
  • Definition 2: OOD and ID Data
  • Definition 3: Ideal Non-member OOD posterior
  • Definition 4: Real-World Non-member OOD
  • Definition 5: Real-World Non-member ID