A New Meta-learning-based Black-box Adversarial Attack: SA-CC

Since deep neural networks are vulnerable to human-imperceptible perturbations with attacking intention, the security research of visual sensors using related detection and recognition algorithms is of value. Then currently black-box adversarial attacks against visual sensors become a popular research direction in this security study for their harsh conditions. However, present black-box attacks have shortcomings in information fully utilization while maintaining a high attack success rate. Our research, based on the newly proposed Simulator Attack, introduces a new attack framework SA-CC by attaching two improving modules for accelerating the whole attack progress. Due to our Cos-Loss and clustering module, the angle information of the output results would be incorporated into the basis of fine-tuning process and simple prior information can be achieved when conducting targeted attack at the beginning. Finally, we perform experiments constantly on the CIFAR-10 dataset. The results obviously exhibit that our attack framework can decrease the total amount of queries considerably and obtain our goal of improving query-efficiency.