GLAMP: Generative Learning for Adversarially-Robust Malware Prediction

We propose a novel Generative Malware Defense strategy. When an antivirus company detects a malware sample <tex-math notation="LaTeX">m</tex-math>, they should: (i) generate a set <tex-math notation="LaTeX">$Var(m)$</tex-math> of several variants of <tex-math notation="LaTeX">m</tex-math> and then (ii) train their malware classifiers on their usual training set augmented with <tex-math notation="LaTeX">$Var(m)$</tex-math>. We believe this leads to a more proactive defense by making the classifiers more robust to future malware developed by the attacker. We formally define the malware generation problem as a non-traditional optimization problem. Our novel GLAMP (Generative Learning for Adversarially-robust Malware Prediction) framework analyzes the complexity of the malware generation problem and includes novel malware variant generation algorithms for (i) that leverage the complexity results. Our experiments show that a sufficiently large percentage of samples generated by GLAMP are able to evade both commercial anti-virus and machine learning classifiers with evasion rates up to 83.81% and 50.54%, respectively. GLAMP then proposes an adversarial training model as well. Our experiments show that GLAMP generates running malware that can evade 11 white box classifiers and 4 commercial (i.e., black box) detectors. Our experiments show GLAMP's best adversarial training engine improves the recall by 16.1% and the F1 score by 2.4%-5.4% depending on the test set used.