Deep learning for image-based mobile malware detection

Current anti-malware technologies in last years demonstrated their evident weaknesses due to the signature-based approach adoption. Many alternative solutions were provided by the current state of art literature, but in general they suffer of a high false positive ratio and are usually ineffective when obfuscation techniques are applied. In this paper we propose a method aimed to discriminate between malicious and legitimate samples in mobile environment and to identify the belonging malware family and the variant inside the family. We obtain gray-scale images directly from executable samples and we gather a set of features from each image to build several classifiers. We experiment the proposed solution on a data-set of 50,000 Android (24,553 malicious among 71 families and 25,447 legitimate) and 230 Apple (115 samples belonging to 10 families) real-world samples, obtaining promising results.