Title: Lethe: Adapter-Augmented Dual-Stream Update for Persistent Knowledge Erasure in Federated Unlearning

Abstract:

Federated unlearning (FU) is designed to eliminate specific knowledge—whether at the client, class, or sample level—from a global model. However, current research typically presumes that the collaborative process concludes immediately following the unlearning procedure, neglecting scenarios where federated training persists on the remaining dataset. In this work, we uncover a significant vulnerability known as "knowledge resurfacing," demonstrating that subsequent training rounds can reactivate previously erased information, thereby restoring its influence within the global model.

To mitigate this issue, we introduce Lethe, an innovative federated unlearning framework that decouples the knowledge targeted for removal from that intended for preservation, thereby guaranteeing lasting erasure even during ongoing training. Lethe operates through a three-stage pipeline: Reshape, Rectify, and Restore. Initially, a temporary adapter is optimized via gradient ascent using the unlearning data to generate amplified updates. These updates serve as corrective signals to direct the layer-wise rectification of subsequent updates across two distinct streams. Following this, the adapter is discarded, and a brief recovery phase is executed on the retained data. Empirical results indicate that Lethe uniformly supports unlearning across all hierarchical levels in federated environments. It demonstrates exceptional persistence, maintaining a resurfacing rate of less than 1% in the majority of cases, despite extensive follow-up training sessions.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC