Classifier-Aware Defense for Visual Recognition in CAV

Investigator(s):

• Yongxin Liu, Embry-Riddle Aeronautical University, ORCID # 0000-0003-4527-8623 (PI)

Project Description:

Figure 1 Classifier-Aware Defense for Visual Recognition in CAV

As in Figure 1, Firstly, we propose a practical comprehensive real-time defense technique based on Compressive Sensing and Generative neural networks (CSG). Compressive Sensing offers a concise signal acquisition paradigm and is used to mitigate the adversarial effects of the perturbed images before feeding them to the classifier for inference. And the generative neural network is adopted to provide fast image reconstruction to meet the timing requirement. Second, we propose a novel incremental training method to improve the comprehensive defense performance of the classifier. Third, we propose a classifier-aware training way to improve the performance of the specific classifier under protection. The goal is achieved by taking the classification loss into account when deriving the training loss of the generative network. The project will produce a set of algorithms to comprehensively defend autonomous driving vision systems against adversarial examples.