Learning a Secure Classifier against Evasion Attack

In security sensitive applications, there is a crafty adversary component which intends to mislead the detection system. The presence of an adversary component conflicts with the stationary data assumption that is a common assumption in most machine learning methods. Since machine learning methods are not inherently adversary-aware, it necessitates to investigate security evaluation of machine learning based detection systems in the adversarial environment. Research in adversarial environment mostly focused on modeling adversarial attacks and evaluating impact of them on learning algorithms, only few studies have devised learning algorithms with improved security. In this paper we propose a secure learning model against evasion attacks on the application of PDF malware detection. The experimental results acknowledge that the proposed method significantly improves the robustness of the learning system against manipulating data and evasion attempts at test time.