English | 中文

This is an official implementation of our paper:Using Highly Compressed Gradients in Federated Learning for Data Reconstruction Attacks. We present a novel data leakage attack algorithm against highly compressed gradients.

To install the requirements, run:

pip install -r requirements.txt

(a) Visualization of HCGLA (Init-Generation) on popular datasets at a 0.1% compression rate with batchsize=1.	(b) Visualization of HCGLA (Init-Generation) on popular facial dataset at a 0.1% compression rate with batchsize=4.

You can download the dataset used in our paper by the following table. Then place these datasets in the corresponding locations refering ProjectStructure.txt.

You can download the models we trained in our experiments by the following table. Then place these models in the corresponding locations refering ProjectStructure.txt.

You need to demonstrate the data reconstruction attack enough times to collect enough noisy images from the reconstruction process and then train a denoising model with the original images and these collected noisy images.

Step 1: To connect enough noising images by running:

python ConnectNoisyimg.py

Step 2: To train denoising model by running:

python Train_DnCNN.py

Step 3: To test denoising model by running:

python Test_DnCNN.py

Then you can see the trained model under the ./models/DenoisingModel/DnCNN folder. You can also download the denoising model we trained by the link：https://pan.baidu.com/s/1EhFDwx8Z4Y3pLySPTz1iLg and its password：vrbs.

When batchsize=1, You can train and test a generator by referring to the Gradinv.html file. When batchsize!=1, you can train a generator by running python Train_batchsize_generator.py, Then you can test a generator by running python Test_batchsize_generator.py. You can download the generators by the link：https://pan.baidu.com/s/1ffmEis1uffoYB69BN_k2pA , password：2p5e. Due to the limited computing power of the lab equipment, we did not train the multi-batch, ResNet generator, but you can also train the ResNet, multi-batch attack scenario generator by yourself if you need.

After prepare all the datasets and models you can launch HCGLA. If you want to launch HCGLA with batchsize = 1, run: python Reconstruct_batchsize1.py. If batchsize != 1, run: python Reconstruct_minibatch.py

You can modify the parameter settings to suit your needs

If you find our work useful in your research, please consider citing:

@ARTICLE{10003066,
  author={Yang, Haomiao and Ge, Mengyu and Xiang, Kunlan and Li, Jingwei},
  journal={IEEE Transactions on Information Forensics and Security}, 
  title={Using Highly Compressed Gradients in Federated Learning for Data Reconstruction Attacks}, 
  year={2023},
  volume={18},
  number={},
  pages={818-830},
  doi={10.1109/TIFS.2022.3227761}}

If you have any questions, please contact me via email [email protected]