An Intelligent Diagnostic Technique for Reliable Operation of Edge AI

With the significant technological evolution of Artificial Intelligence (AI) and improved performance of AI processing units, AI algorithms such as Deep Neural Networks (DNNs) are widely used on the edge side in various fields, including robotics, medicine, and mobility. Nowadays, AI on the edge side, i.e., edge AI, is even used for safety-critical applications such as automated driving applications. On the other hand, several challenges still exist to ensure the reliable operation of edge AI, as hardware faults and security attacks can lead to unexpected changes in AI parameters and consecutive incorrect output. However, there is no established technique that confirms that edge AI works correctly during the operation. Therefore, an intelligent diagnostic technique that can detect unexpected changes during AI inference is proposed in this paper. Conventional Convolutional Neural Network (CNN) models are considered in this study, and experimental results show that the proposed technique can effectively diagnose approximately 97% cases of unexpected change in AI parameters. Experimental results show the effectiveness of the proposed technique, which can be applied to various AI processing units without any modification of the AI model and thus can contribute to ensure the reliable operation of edge AI in various fields.