Introduction to the Special section on artificial intelligence security: Adversarial attack and defense

Artificial intelligence (AI) has been widely adopted in various applications such as face detection, speech recognition, machine learning, etc. Due to the lack of theoretical explanation, recent works show that AI is vulnerable to adversarial attacks, especially deep neural networks could be easily fooled by adversarial examples that are in the form of subtle perturbations to the inputs. The intrinsic vulnerability of AI might incur severe security problems in areas like automatic driving, face payment, voice command control, etc. Adversarial learning is one typical defense method, which can migrate such security risk of AI by training with generated adversarial examples. However, this method cannot defend the growing number of adversarial attacks. The special section of “Artificial Intelligence Security: Adversarial Attack and Defense” focused on the state-of-art adversarial attack and defense methods, and explored how these security problems could affect other areas such as cyberspace security, and internet-of-things.