HF-GradInv

Implementation of AAAI 2024 paper "High-Fidelity Gradient Inversion in Distributed Learning"

Requirements

I have tested on:

PyTorch 1.13.0
CUDA 11.0

The Simplest Implementation

If you want to test the gradient inversion attacks:

python main_attack.py

If you want to test with duplicated labels:

python main_attack_duplicate_labels.py

Test images, reconstructions, as well as the auxiliary data we used in our paper are available in folder "custom_data".

It is easy to test on more settings, just need to adjust the corresponding parameters as well as the training data in "custom_data".

REFERENCES

https://github.com/JonasGeiping/breaching