Explaining Vulnerabilities of Deep Learning to Adversarial Malware Binaries

摘要: Recent work has shown that deep-learning algorithms for malware detection are also susceptible to adversarial examples, i.e., carefully-crafted perturbations input enable misleading classification. Although this questioned their suitability task, it is not yet clear why such easily fooled in particular application domain. In work, we take a first step tackle issue by leveraging explainable machine-learning developed interpret the black-box decisions of deep neural networks. particular, use an technique known as feature attribution identify most influential features contributing each decision, and adapt provide meaningful explanations classification binaries. case, find recently-proposed convolutional network does learn any characteristic from data text sections executable files, but rather tends discriminate between benign samples based on characteristics found file header. Based finding, propose novel attack algorithm generates binaries only changing few tens bytes With respect other state-of-the-art algorithms, our require injecting padding at end file, much more efficient, requires manipulating fewer bytes.