Development of hybrid multi-stage reasoning method for malicious behaviors identification in nuclear security

To prevent nuclear facilities and nuclear materials from being destroyed and unauthorized accessed by threateners, top priority should be given to the identification of such human malicious behaviors in a reliable and efficient way. Nowadays, the research on human malicious behaviors identification is obviously promoted based on the rapid development of deep learning techniques. The problem of human malicious behaviors identification turns out to be complex and tricky, while both the spatial and temporal dependencies of human actions should be comprehensively analyzed. In our previous works, three novel reasoning methods, which are data-based, language-based and graph-based reasoning methods, are developed for this problem achieving promising results. In this research, to further advance the performance, we first conduct a qualitative analysis for these three reasoning methods, especially for the aspects of spatial and temporal dependencies. Then, hybrid multi-stage reasoning methods are developed, containing two-stage reasoning method and three-stage reasoning method. In these reasoning methods, the graph-based reasoning method, which could comprehensively analyze both the spatial and temporal dependencies of human action sequences, is operated at first. Then, the other two reasoning methods focusing more on spatial dependencies are conducted. For evaluating the performances of the advanced methods, a self-collected nuclear security dataset is applied for experiments. This dataset is comprised of five typical malicious behaviors in nuclear security which are fence climbing, wire net cutting, weapon holding, armed boundary sabotage and nuclear material theft. The results show that the proposed two-stage hybrid reasoning method, which operates graph-based reasoning method first and then operates language-based reasoning method, achieves state-of-the-art results. Especially, it achieves optimal accuracy value of 0.8472 and optimal F1-score of 0.8534, while obtaining relatively high precision value of 0.7805 and high recall value of 0.9412.