Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Federated Domain Unlearning: Verification Methodologies and Challenges

Kahou Tam, Kewei Xu, Li Li, Huazhu Fu

TL;DR

This work addresses federated unlearning under domain heterogeneity, motivated by RTBF, and shows that existing FL unlearning methods fail to selectively erase domain-specific information without harming non-targeted domains. It provides an empirical study across multi-domain benchmarks, revealing that deep layers drive forgetting and that current methods often cause collateral degradation. To enable reliable domain-specific validation, the authors propose a four-stage verification framework that uses domain-representative samples and adversarial markers to assess unlearning efficacy while constraining global performance. The framework is demonstrated on DomainNet, Domain-Digits, and Office-Caltech, offering practical insights and a concrete path toward domain-centric unlearning in FL.

Abstract

Federated Learning (FL) has evolved as a powerful tool for collaborative model training across multiple entities, ensuring data privacy in sensitive sectors such as healthcare and finance. However, the introduction of the Right to Be Forgotten (RTBF) poses new challenges, necessitating federated unlearning to delete data without full model retraining. Traditional FL unlearning methods, not originally designed with domain specificity in mind, inadequately address the complexities of multi-domain scenarios, often affecting the accuracy of models in non-targeted domains or leading to uniform forgetting across all domains. Our work presents the first comprehensive empirical study on Federated Domain Unlearning, analyzing the characteristics and challenges of current techniques in multi-domain contexts. We uncover that these methods falter, particularly because they neglect the nuanced influences of domain-specific data, which can lead to significant performance degradation and inaccurate model behavior. Our findings reveal that unlearning disproportionately affects the model's deeper layers, erasing critical representational subspaces acquired during earlier training phases. In response, we propose novel evaluation methodologies tailored for Federated Domain Unlearning, aiming to accurately assess and verify domain-specific data erasure without compromising the model's overall integrity and performance. This investigation not only highlights the urgent need for domain-centric unlearning strategies in FL but also sets a new precedent for evaluating and implementing these techniques effectively.

Towards Federated Domain Unlearning: Verification Methodologies and Challenges

TL;DR

This work addresses federated unlearning under domain heterogeneity, motivated by RTBF, and shows that existing FL unlearning methods fail to selectively erase domain-specific information without harming non-targeted domains. It provides an empirical study across multi-domain benchmarks, revealing that deep layers drive forgetting and that current methods often cause collateral degradation. To enable reliable domain-specific validation, the authors propose a four-stage verification framework that uses domain-representative samples and adversarial markers to assess unlearning efficacy while constraining global performance. The framework is demonstrated on DomainNet, Domain-Digits, and Office-Caltech, offering practical insights and a concrete path toward domain-centric unlearning in FL.

Abstract

Federated Learning (FL) has evolved as a powerful tool for collaborative model training across multiple entities, ensuring data privacy in sensitive sectors such as healthcare and finance. However, the introduction of the Right to Be Forgotten (RTBF) poses new challenges, necessitating federated unlearning to delete data without full model retraining. Traditional FL unlearning methods, not originally designed with domain specificity in mind, inadequately address the complexities of multi-domain scenarios, often affecting the accuracy of models in non-targeted domains or leading to uniform forgetting across all domains. Our work presents the first comprehensive empirical study on Federated Domain Unlearning, analyzing the characteristics and challenges of current techniques in multi-domain contexts. We uncover that these methods falter, particularly because they neglect the nuanced influences of domain-specific data, which can lead to significant performance degradation and inaccurate model behavior. Our findings reveal that unlearning disproportionately affects the model's deeper layers, erasing critical representational subspaces acquired during earlier training phases. In response, we propose novel evaluation methodologies tailored for Federated Domain Unlearning, aiming to accurately assess and verify domain-specific data erasure without compromising the model's overall integrity and performance. This investigation not only highlights the urgent need for domain-centric unlearning strategies in FL but also sets a new precedent for evaluating and implementing these techniques effectively.
Paper Structure (31 sections, 14 equations, 13 figures, 15 tables)

This paper contains 31 sections, 14 equations, 13 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Federated Learning with Domain Heterogeneity
  4. Federated Unlearning
  5. Federated Domain Unlearning
  6. Comparison with Prior Federated Unlearning
  7. Empirical Study
  8. Setup
  9. Effectiveness of Existing Methods in Federated Domain Unlearning
  10. Federated Domain Unlearning and Hidden Layer Representations
  11. Feature Reuse
  12. New validation methods for Federated domain unlearning
  13. Current Verification Methods and Their Limitations
  14. Methodology Overview
  15. Domain Representative Sample Selection
  16. ...and 16 more sections

Figures (13)

  • Figure 1: The workflow of federated unlearning.
  • Figure 2: CKA Analysis of Layer Representations Before and After Unlearning the Target Domain in DomainNet. We select three domains to display: (a) Infograp, (b) Painting, and (c) Sketch. More detail results are shown in supplementary materials.
  • Figure 3: Comparative CKA Analysis of Layer Representations in Unlearned and Remaining Domains in DomainNet. We report the results of the methods Repaid-Retrain and FedEraser, which unlearn the target domain but also impact the remaining domain's integrity.
  • Figure 4: Comparative Analysis of Subspace Similarity in Feature Extractors Before and After Unlearning in the Target Domain of DomainNet.
  • Figure 5: Validation Results of Our Verification Method Under Different Parameters.
  • ...and 8 more figures