EFFACE: Top-1 Compression Suffices for Federated Unlearning with the Help of Adaptive Error Feedback

This is the official code repository for the ICASSP 2026 paper Top-1 Compression Suffices for Federated Unlearning with the Help of Adaptive Error Feedback

🐱‍💻 Code

See EFFACE/LogisticRegression for the details of toy example and unlearning trajectory visualization, and EFFACE/ImageClassification for the details of image classification experiments.

🔍 Clarification on the Selection Mechanism

The selection mechanism presented in the ICASSP paper was optimized for empirical performance. However, our subsequent theoretical analysis indicates that incorporating a correction term is necessary to strictly guarantee convergence bounds.

Formulation Comparison

Version	Formula
Original (Used in ICASSP Conference Paper Experiments)
Corrected (Used in the Upcoming Extended Version and Current Repo Implementation)

Implementation Update: We have updated the corresponding code to implement the theoretically corrected version. Our empirical verification confirms that this update yields performance comparable to the original implementation, ensuring both theoretical rigor and practical effectiveness.

🚀 Upcoming Extended Version

We are preparing a comprehensive extension of this work that will include:

1. Convergence Analysis: Detailed analysis of the convergence order for EFFACE, including the selection of the error compensation strength coefficient $\eta$.
2. More mechanisms designed for Federated Unlearning: Empirical mechanisms designed for federated unlearning to improve the unlearning performance.

This extended manuscript is currently in preparation and will be made available soon. We appreciate your interest and patience.

📖 Citation

If you find this work useful, please consider citing:

@inproceedings{xiao2026efface,
  title={Top-1 Compression Suffices for Federated Unlearning with the Help of Adaptive Error Feedback},
  author={Xiao, Boxu and Liu, Sijia and Ling, Qing},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={2171--2175},
  year={2026},
  organization={IEEE}
}