FTCLNet: Convolutional LSTM with Fourier Transform for Vulnerability Detection

As software vulnerabilities become increasingly serious, it is necessary to detect them efficiently and accurately. However, vulnerabilities are diverse and context sensitive. Previous solutions either rely on features defined by experts, or use only recurrent neural networks on code sequence. It is difficult to extract complex features of vulnerabilities in traditional code space. This article proposes a deep convolutional LSTM neural network with Fourier transform for vulnerability detection. The discrete Fourier transform method convert code space into frequency domain, which significantly helps deep models learn remarkable patterns. This article combines convolutional neural network (CNN) with long short term memory (LSTM) network to extract local and global features in frequency domain, and utilize attention mechanism to decide the weight of each element in code space. Besides, this method rewrite the source code and convert them to vectors without guidance from the specified domain knowledge. Experiments on Buffer Error dataset (CWE-119) and Resource Management Error dataset (CWE-399) show that this new method achieves a significantly improved results.