Vax-a-Net: Training-Time Defence Against Adversarial Patch Attacks

We present Vax-a-Net; a technique for immunizing convolutional neural networks (CNNs) against adversarial patch attacks (APAs). APAs insert visually overt, local regions (patches) into an image to induce misclassification. introduce conditional Generative Adversarial Network (GAN) architecture that simultaneously learns synthesise patches use in APAs, whilst exploiting those adapt pre-trained target CNN reduce its susceptibility them. This approach enables resilience be conferred models, which would impractical with conventional training due the slow convergence of APA methods. demonstrate transferability this protection defend existing and show efficacy across several contemporary architectures.