AWA: Approximate and Weighted Attack in
Federated Learning

Yongcun Song, Ziqi Wang, and Enrique Zuazua

Overview

This repository implements Approximate and Weighted Attack (AWA) for reconstructing private client data from federated learning updates. The method combines update approximation with layer-wise objective weighting, and the weighting parameters can be tuned via Bayesian optimization.

Core components:

• Reconstruction entry point: main_AWA.py
• Bayesian optimization entry point: main_BO.py
• Attack implementation: attacks/
• Default configuration: utils/config.yaml

Usage

Run a single reconstruction experiment:

python main_AWA.py --cfg utils/config.yaml

Run Bayesian optimization over the weighting parameters:

python main_BO.py --cfg utils/config.yaml --trials 100

Citation

If this code is useful for your research, please cite the paper:

@article{swz2026AWA,
 author={Song, Yongcun and Wang, Ziqi and Zuazua, Enrique},
  journal={IEEE Transactions on Big Data}, 
  title={Approximate and Weighted Data Reconstruction Attack in Federated Learning}, 
  year={2026},
  doi={10.1109/TBDATA.2026.3695416}
}

Acknowledgments

Alphabetical authorship according to mathematical tradition. Funded by the European Union's Horizon Europe MSCA project ModConFlex (grant number 101073558)