PROPOSED METHOD TO REMOVE ADVERSARIAL PERTURBATION USING GENERATIVE MODEL BASED ON DEEP LEARNING

Abstract

With the rapid advancement of information technology, artificial intelligence has found extensive applications in various fields, including object recognition, facial recognition, autonomous vehicle operation, and healthcare. However, deep neural networks, which serve as the foundation of many artificial intelligence systems, are highly vulnerable to adversarial examples. These adversarial examples are crafted by introducing subtle and imperceptible perturbations into clean images, effectively deceiving artificial intelligence models and exposing critical weaknesses. Addressing this challenge, the authors propose a new method to remove adversarial perturbation present in the images. This method employs a data generator that learns features directly from the input images, enabling the reconstruction of clean (adversarial perturbations has been removed). The research results demonstrate that this method not only effectively mitigates noise in individual adversarial examples but also counters attacks utilizing adversarial images. This approach opens a new pathway to enhance the accuracy and security of artificial intelligence applications in practice.