Deep Learning (DL) techniques were successfully applied to solve challenging problems in the field of Natural Language Processing (NLP). Since source code and natural text share several similarities, it was possible to adapt text classification techniques such as word embedding to propose DL-based Automatic Vulnerabilities Prediction (AVP) approaches. Although the obtained results were interesting, they were not good enough as it was obtained in LNP. In this paper, we propose an improved DL-based AVP approach based on the technique of character n-gram embedding. We evaluate the proposed approach for 4 types of vulnerabilities using a large c/c++ open-source codebase. The results show that our approach can get very excellent performances and outperform the obtained performances by the previous approaches.
Key words: Software Security, Vulnerability Detection, Deep Features Learning, Character N-gram Embedding.