Abstract:

Deep learning models have performed well on many AI tasks, but elaborate adversarial samples can trick well-trained models into making false judgments. The success of the adversarial attack calls into question the usability of the AI system. In order to improve the security and robustness, the paper follow the security development lifecycle and propose a security evaluation and defense reinforcement scheme for the AI system. The scheme improves the system's ability to resist attacks and helps developers build a more secure AI system through measures such as accurate detection and interception of adversarial attacks, scientific evaluation of the model's robustness, and real-time monitoring of new adversarial attacks.

Key words: deep learning, adversarial attack, security development lifecycle