Malware Detection with Malware Images using Deep Learning Techniques

Driven by economic benefits, the number of malware attacks is increasing significantly on a daily basis. Malware Detection Systems (MDS) is the first line of defense against malicious attacks, thus it is important for malware detection systems to accurately and efficiently detect malware. Traditional MDS typically utilizes traditional machine learning algorithms that require feature selection and extraction, which are time-consuming and error-prone. Conventional deep learning based approaches typically use Recurrent Neural Network (RNN) which can be vulnerable to redundant API injection. Thus, we investigate the effectiveness of Convolutional Neural Networks (CNN) against redundant API injection. We designed a malware detection system that transforms malware files into image representations and classifies the image representation with CNN. The CNN is implemented with spatial pyramid pooling layers (SPP) to deal with varying size input. We evaluate the effectiveness of SPP and image color space (greyscale/RGB) by measuring the performance of our system on both unaltered data and adversarial data with redundant API injected. Results show that naive SPP implementation is impractical due to memory constraints and greyscale imaging is effective against redundant API injection.