Search Papers | Poster Sessions | All Posters
Poster C104 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
How Predictive Coding Rescues Feed-Forward Networks on Adversarial Attacks
Ehsan Ganjidoost1,2,3 (), Jeff Orchard1,2,3; 1University of Waterloo, 2Cheriton School of Computer Science, 3Neurocognitive Computing Lab
This study introduces Predictive Coding Networks (PCnets) as a defence mechanism against adversarial attacks on neural network classifiers. By integrating PCnets into Feed-Forward Networks (FFnets), we enhance their resilience to adversarial perturbations. Using MNIST, we experimentally demonstrate the effectiveness of PCNets in identifying and mitigating adversarial examples generated to attack a fully-connected network, and a CNN. Leveraging the generative nature of PCnets, the defence mechanism effectively counters adversarial efforts, reverting perturbed images closer to their original forms. This innovative approach presents a promising solution for improving the security and reliability of neural network classifiers amidst the rising threat of adversarial attacks.
Keywords: Predictive Coding Perturbation Attack Biological Plausibility Adversarial Attack