Fig. 1
From: Iterative and mixed-spaces image gradient inversion attack in federated learning
An overview of our attack framework, including two recovery algorithms, separately for image (Left, blue dashed box) and label (right, black dashed lines). For image reconstruction, black solid lines denote the forwarding process, light dashed lines denote the back propagation, and the red dashed box contains the real optimized variables. In each attack round, the attacker first collects shared gradients from the server side, then recovers the label, and last inverts the image through optimization