Research of Black-Box Adversarial Attack Detection Based on GAN

The abnormal behavior detector based on machine learning is vulnerable to adversarial attacks. Attackers achieve adversarial attacks with constructing counter samples on the detection algorithm, and get the expected output for the target model, by changing the attack behavior pattern and API call sequence. In order to improve the performance of black-box attack, this paper can generate adversarial sample to deceive discriminator by training the generator and discriminator through the black-box adversarial attack network based on GAN. Experiments on real world API attacks and benchmark data sets have proved that the anti-attack anomaly behavior detection system can well detect malware and anomaly behavior.