人工智能系统安全与隐私风险

Human society is witnessing a wave of artificial intelligence(AI) driven by deep learning techniques, bringing a technological revolution for human production and life. In some specific fields, AI has achieved or even surpassed human-level performance. However, most previous machine learning theories have not considered the open and even adversarial environments, and the security and privacy issues are gradually rising. Besides of insecure code implementations, biased models, adversarial examples, sensor spoofing can also lead to security risks which are hard to be discovered by traditional security analysis tools. This paper reviews previous works on AI system security and privacy, revealing potential security and privacy risks. Firstly, we introduce a threat model of AI systems, including attack surfaces, attack capabilities and attack goals. Secondly, we analyze security risks and counter measures in terms of four critical components in AI systems: data input(sensor), data preprocessing, machine learning model and output. Finally, we discuss future research trends on the security of AI systems. The aim of this paper is to arise the attention of the computer security society and the AI society on security and privacy of AI systems, and so that they can work together to unlock AI’s potential to build a bright future.