Abstract
Industrial cyber-physical systems (CPSs) play a more important role in today’s world with the fast development of Industry 4.0. Based on connecting the cyber world to the physical world, cyber-physical systems are widely used in various applications such as delivery, transportation, and medical devices to achieve the Internet of Things (IoT). However, CPSs are complicated, especially those with "smart" systems, for example CPSs with artificial intelligence (AI) algorithms. Considered "soft” algorithms, AI-based algorithms operate very differently from traditional "hard” algorithms and are increasingly used in CPSs. We can quickly generate an AI model with good performance for different purposes with enough training data, but it is hard to explain the prediction of AI models [1], and thus it is hard to defend an AI-driven system. There are different types of AI models, and deep learning models are one of the main adoptions for AI-driven systems. In general, deep learning models are implemented with deep neural networks, which have shown outstanding performance in various areas, such as computer vision (CV), natural language processing (NLP), audio processing, Etc. However, neural networks are vulnerable to adversarial attacks [2]. A small designed change of input data can lead to a totally different output. Given that CPSs in the industry are highly safety-critical, this work focuses on defending industrial CPSs against adversarial attacks. To defend the industrial CPSs, we firstly propose an adversarial attack method to evaluate a real-world water treatment system. This work also analyzes the difference between adversarial attacks on AI models and AI-driven CPSs. There is less work focusing on implementing adversarial attacks on industrial AI-driven CPSs. Furthermore, we improve the adversarial attack method to make the attack more applicable for industrial systems by generating physical adversarial patches to interfere with deep learning models. In this work, we evaluate our attack on a robotic arm, which is the essential intelligent manufacturing system common in the CPS of Industry 4.0. Afterward, we propose a defending method to improve the AI model robustness by selecting boundary data for model training to defend against such attacks. Finally, we design the approach considering the requirements of the industrial process and evaluate the defense method on the industrial robotic arm