AEGuard: Image Feature-Based Independent Adversarial Example Detection Model

Abstract

With the rapid development of image processing technology, image recognition systems based on massive image data are being developed and deployed. The wrong decision regarding an image recognition system for security-sensitive systems can cause serious problems such as personal accidents and property damage. Furthermore, adversarial attacks, which are security attacks that cause malfunctions in image recognition systems by inserting adversarial noise, have emerged and evolved. Several studies have been conducted to prevent adversarial attacks. However, existing mechanisms have low classification accuracy and low detection accuracy for adversarial examples with small adversarial noise. This paper proposes an adversarial example detection mechanism based on image feature extraction and a deep neural network (DNN) model. The proposed system achieves versatility and independence by detecting adversarial examples based on image features, such as edge noise and discrete cosine transform (DCT) bias, which adversarial examples have in common. The proposed system shows relatively higher detection accuracy than existing mechanisms for various types and amounts of adversarial noise and different sharpness of adversarial examples because the proposed system detects them depending on the characteristics of each type of adversarial example.