FingerprintAttack

This repository is dedicated to Fingerprint Attack: Client De-Anonymization in Federated Learning (accepted to ECAI 2023). We demonstrate a new fingerprinting attack on the shuffle module in federated learning (FL).

Step 0: Get Started

• Install PyTorch: python=3.10, pytorch-cuda=11.7, pytorch, torchvision, torchaudio, sentencepiece, transformers, datasets, sklearn etc.;
• Install local FLSim (for simulating FL): 1. download/clone FLSim code; 2. add some modifications to save gradient updates (more details in next section); 3. install local FLSim;
• Install opacus_lab (for refactoring GPT2): download/clone opacus_lab code and add dir_path to lm_exp.py.

Step 1: Run FL Simulator

python lm_exp.py --config-file configs/news_fed_config.json 
python lm_exp.py --config-file configs/dial_fed_config.json  

Note that: the gradients of the models can be recorded by adding the following code to '_update_clients' (FLSim: trainers/sync_trainer.py) and some minor corresponding changes.

import os
path = os.path.join(self.cfg.client.store_models_dir, '{}_e{}.pt'.format(client.name, epoch))
torch.save([(name, w.grad.data.cpu()) for name, w in client_delta.model.named_parameters()], path)

Step 2: Run Analysis (Clustering and Alignment)

python analyse_cluster.py --cluster greedy_match

python analyse_cluster.py --cluster spectral

python analyse_cluster.py --cluster kmean

Citation

This work is to appear in ECAI 2023:

@inproceedings{
  xu2023fingerprint,
  title={Fingerprint Attack: Client De-Anonymization in Federated Learning},
  author={Qiongkai Xu and Trevor Cohn and Olga Ohrimenko},
  booktitle={The 26th European Conference on Artificial Intelligence},
  year={2023}
}