On Knowledge Editing in Federated Learning: Perspectives, Challenges, and Future Directions

Leijie Wu, Song Guo, Junxiao Wang, Zicong Hong, Jie Zhang, Jingren ZhouarXiv 2023 · DOI · arXiv

Abstract

As Federated Learning (FL) transitions from research settings to dynamic real-world applications, models must adapt to both acquiring new skills and "forgetting" specific data for privacy or efficiency. This paper introduces Federated Editable Learning (FEL), a unified framework that categorizes state-of-the-art techniques in knowledge augmentation (continual learning) and knowledge removal (unlearning), while emphasizing the critical role of verification and auditing in the model lifecycle.

Introduction

Standard Federated Learning typically assumes a static environment with fixed data distributions and tasks. However, real-world deployments face two primary "dynamics":

  1. Data Dynamics: Local clients continuously receive new data or tasks. This often leads to catastrophic forgetting, where the global model loses previously acquired knowledge while attempting to integrate new information.
  2. Device Dynamics: Clients frequently join or leave the network. Privacy regulations, such as the GDPR, grant users the "right to be forgotten," necessitating methods to remove a specific client's influence from the global model without the prohibitive cost of retraining from scratch—a process known as Federated Unlearning (FU).

This paper addresses the gap in current literature by providing a holistic view of "Knowledge Editing"—the ability to precisely add or subtract information from a distributed model.

The Federated Editable Learning (FEL) Paradigm

The authors propose the Federated Editable Learning (FEL) framework, which structures the field into three main pillars: knowledge increase, knowledge removal, and verification.

The Federated Editable Learning (FEL) lifecycle, illustrating the flow from data joining (learning) to data exiting (unlearning) and the necessary verification steps.

1. Federated Increasing Learning (FIL)

FIL focuses on expanding the model's knowledge base through two primary sub-fields:

  • Federated Continual Learning (FCL): Employs techniques to mitigate catastrophic forgetting, including regularization-based approaches (penalizing changes to critical weights) and exemplar-based methods (retaining a subset of old samples).
  • Federated Few-shot Learning (FFsL): Utilizes meta-learning and adversarial training to enable models to generalize to new tasks using very limited local data.

2. Federated Unlearning (FU)

FU focuses on the precise removal of specific knowledge, categorized by the "exactness" of the operation:

  • Exact Unlearning: Involves retraining or using modular architectures like LegoNet, which employs isolated "adapters" that can be swapped or deleted to remove specific influences.
  • Approximate Unlearning: Offers faster "scrubbing" methods, such as Gradient Recovery (reversing model states using historical updates) or Channel Pruning (identifying and removing neurons associated with the forgotten data).
Taxonomy of Federated Unlearning methods, categorized by exactness, scenario, and the level of the unlearning request (Client, Class, or Sample).

3. Verification & Auditing

The authors argue that a model "edit" is only valid if it can be verified. They review methods such as Trusted Execution Environments (TEEs) and Proof of Learning to ensure honest training, alongside Membership Inference Attacks to confirm that "unlearned" data has indeed been removed.

Key Insights and Future Directions

The survey highlights several critical perspectives for the future of FL:

  • The Capacity Limit: Models have a finite "knowledge upper limit." Unlearning serves as a management tool to remove obsolete data, freeing capacity for new, relevant knowledge.
  • Knowledge Disentanglement: Currently, knowledge in neural networks is highly entangled. Removing specific data points often inadvertently degrades performance on unrelated tasks. Moving toward modular, "Lego-like" architectures is a promising path for efficient editing.
  • Trustworthy Communities: Integrating Blockchain with FL could create a transparent, tamper-proof ledger of all model edits, ensuring accountability in decentralized AI systems.

Conclusion

Federated Editable Learning represents a shift toward more flexible and compliant distributed AI. By unifying learning and unlearning under a single framework of "Knowledge Editing," researchers can better address the complexities of dynamic real-world data and privacy requirements.

Summary updated

Abstract

As Federated Learning (FL) has gained increasing attention, it has become widely acknowledged that straightforwardly applying stochastic gradient descent (SGD) on the overall framework when learning over a sequence of tasks results in the phenomenon known as ``catastrophic forgetting''. Consequently, much FL research has centered on devising federated increasing learning methods to alleviate forgetting while augmenting knowledge. On the other hand, forgetting is not always detrimental. The selective amnesia, also known as federated unlearning, which entails the elimination of specific knowledge, can address privacy concerns and create additional ``space'' for acquiring new knowledge. However, there is a scarcity of extensive surveys that encompass recent advancements and provide a thorough examination of this issue. In this manuscript, we present an extensive survey on the topic of knowledge editing (augmentation/removal) in Federated Learning, with the goal of summarizing the state-of-the-art research and expanding the perspective for various domains. Initially, we introduce an integrated paradigm, referred to as Federated Editable Learning (FEL), by reevaluating the entire lifecycle of FL. Secondly, we provide a comprehensive overview of existing methods, evaluate their position within the proposed paradigm, and emphasize the current challenges they face. Lastly, we explore potential avenues for future research and identify unresolved issues.
BibTeX
@misc{wu2023knowledge,
  author = {Leijie Wu and Song Guo and Junxiao Wang and Zicong Hong and Jie Zhang and Jingren Zhou},
  title = {{On Knowledge Editing in Federated Learning: Perspectives, Challenges, and Future Directions}},
  year = {2023},
  eprint = {2306.01431},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.01431},
}
Figure