Fast and Accurate Deep Leakage from Gradients Based on Wasserstein Distance

Figure 1

Overview of our WDLG algorithm. Normal participant inputs training sample to obtain the original gradient , and malicious attacker randomly initializes the virtual data to obtain the virtual gradient and minimizes the Wasserstein distance between the original gradient and the virtual gradient . When the iterative optimization is completed, the virtual data randomly initialized by a malicious attacker converges to with a preset threshold, thereby obtaining the training sample from the normal participant.