Fig. 5: Visual demonstration of enhanced privacy protection through gradient inversion resistance.
From: Concept transfer of synaptic diversity from biological to artificial neural networks
Comparison of original training images (left column) with reconstructed images from gradient information using baseline AlexNet (middle column) and biomod AlexNet (right column) after 100 epochs of training. Biomod significantly degrades reconstruction quality, evidenced by increased pixelation and loss of recognizable features, demonstrating enhanced protection against privacy attacks in distributed learning scenarios.
