GGI: Generative Gradient Inversion Attack in Federated Learning

Recent studies have indicated that the private client images in federated learning can be restored through the shared global model by gradient based inversion attacks. These attacks often employ gradient matching strategy to align the gradients of the reconstructed dummy data with the real data, thereby bringing the dummy images look similar to the actual private images. However, due to the inherent characteristic of high dimensional search space, directly optimizing the dummy images is easy to cause reconstruction failure. In addition, the quality of restored images is significantly sensitive to the data initialization method. Therefore, in this paper, we would like to introduce a novel generative gradient inversion, named GGI, attack algorithm where the dummy images are generated on low-dimensional latent vectors through a pre-trained generator. Compared to the common entire data reconstruction, optimizing the random vectors dramatically reduce the dimension of search space. Our empirical results demonstrate that our proposed method can effectively enhance the quality of restored images.