EITGAN: A Transformation-based Network for recovering adversarial examples

• DOI: 10.3934/era.2023335
• Publication Date: 2023-10-16T10:40:05Z
• Authors: Junjie Zhao; Junfeng Wu; James Msughter Adeke; Guangjie Liu; Yuewei Dai
• Keywords: Generative adversarial network; Deep Neural Networks; Retraining

<abstract><p>Adversarial examples have been shown to easily mislead neural networks, and many strategies proposed defend them. To address the problem that most transformation-based defense will degrade accuracy of clean images, we an Enhanced Image Transformation Generative Adversarial Network (EITGAN). Positive perturbations were employed in EITGAN counteract adversarial effects while enhancing classified performance samples. We also used image super-resolution method mitigate effect perturbations. The does not require modification or retraining classifier. Extensive experiments demonstrated enhanced samples generated by effectively defended against attacks without compromising human visual recognition, their classification was superior images.</p></abstract>