Detecting adversarial examples in object detection: positive and negative representation approach

The susceptibility of deep neural networks to adversarial examples remains a significant hurdle, particularly within the realm of object detection. Given the unique nature of tasks in the field of object detection, many existing adversarial example detection methods designed for image classification are not directly applicable. To address this challenge, we propose the object detection positive and negative detector. This framework identifies adversarial instances by capitalizing on the disparities in predictive outcomes between positive and negative object representations. It integrates a positive and negative object detector with a cascade consistency classification algorithm, facilitating high-fidelity detection and rigorous verification of predictive consistency. Our approach was rigorously assessed using a comprehensive set of 10 adversarial examples across a diverse array of object detection models, encompassing three single-stage and two two-stage architectures. The experimental outcomes reveal that our method attains an average detection rate of 0.86, with a peak performance of 0.99, substantiating its efficacy. Furthermore, the consistent detection rate of ∼0.86 across various models and adversarial examples underscores its robust stability and reliability.