Federated Unlearning: Concept & Challenges

Published:

Federated Unlearning: Concept & Challenges

This post is based on the talk Learning and Unlearning Your Data in Federated Settings (PEPR ‘24, USENIX).
이 글은 Learning and Unlearning Your Data in Federated Settings 발표를 기반으로 정리한 내용이다.

Overview

Summary:
An overview of federated unlearning and its key challenges in balancing privacy, efficiency, and model utility.

🔑 Research Question:

  • Can federated learning systems support safe and efficient data deletion without full retraining?

⚙️ Conceptual Approaches:

  • Passive Unlearning:
    • Server-only (leveraging stored updates)
    • Client-aided (clients assist with gradient/history)
  • Active Unlearning:
    • Server and clients collaboratively remove the influence of target data
  • Levels of Unlearning:
    • Record-level, class-level, or client-level

📊 Key Insights:

  • Retraining is reliable but computationally expensive
  • Approximate unlearning improves efficiency but weakens guarantees
  • Privacy, consistency, and efficiency must be balanced
  • Lack of formal verification remains a core challenge

⚠️ Limitations & Open Challenges:

  • Verifiability: proving that unlearning actually occurred
  • Dynamic participation: handling clients joining/leaving
  • Fairness and explainability remain underexplored
  • New privacy risks may arise during unlearning

💡 Insight:
Federated unlearning introduces a fundamental tension between data deletion guarantees and system efficiency, suggesting that future work must integrate both cryptographic guarantees and system-level design.


Slides:
PDF (Korean) Download