Empowering convolutional networks for malware classification and analysis

摘要: Performing large-scale malware classification is increasingly becoming a critical step in analytics as the number and variety of samples rapidly growing. Statistical machine learning constitutes an appealing method to cope with this increase it can use mathematical tools extract information out datasets produce interpretable models. This has motivated surge scientific work developing methods for detection malicious executables. However, optimal extracting most informative features different families, final goal classification, yet be found. Fortunately, neural networks have evolved state that they surpass limitations other terms hierarchical feature extraction. Consequently, now offer superior accuracy many domains such computer vision natural language processing. In paper, we transfer performance improvements achieved area model execution sequences disassembled binaries. We implement network consists convolutional feedforward constructs. architecture embodies extraction approach combines convolution n-grams instructions plain vectorization derived from headers Portable Executable (PE) files. Our evaluation results demonstrate our outperforms baseline methods, simple Feedforward Neural Networks Support Vector Machines, achieve 93% on precision recall, even case obfuscations data.