Adversarially Robust Malware Detection Using Monotonic Classification
Inigo Incer, Michael Theodorides, Sadia Afroz, David Wagner · 2018
We propose monotonic classification with selection of monotonic features as a defense against evasion attacks on classifiers for malware detection. The monotonicity property of our classifier ensures that an adversary will not be able to evade the classifier by adding more features. We train and test our classifier on over one million executables collected from VirusTotal. Our secure classifier has 62% temporal detection rate at a 1% false positive rate. In comparison with a regular classifier with unrestricted features, the secure malware classifier results in a drop of approximately 13% in detection rate. Since this degradation in performance is a result of using a classifier that cannot be evaded, we interpret this performance hit as the cost of security in classifying malware.