Defense against adversarial malware using robust classifier: DAM-ROC

Malware authors focus on deceiving and evading Anti Malware Engines (AME). Evasion attacks take in malware samples and modify those samples to by-pass ml based AME. Existing learning based anti-malware solutions are either too limited or insufficient to neutralize the threats arising from evasion attacks. This has necessitated a more comprehensive and robust solution. This research attempts to develop a secure learning framework entitled, damroc. The objective is to shield anti-malware entities against evasion attacks by making use of an adaptive adversarial training framework with novel retraining sample selector, (DAM-ROC OR) for dnn based learners. Usage of bnn along with possible quantification of predictive uncertainties is adapted. This generic framework, DAM-ROC is evaluated on benchmarked Android and Windows datasets to explore necessary trade-off between performance and robustness. DAM-ROC models are retrained to defend against gradient attacks like rBIMk, dBIMk, GRAMS and JSMA. Empirical results show that proposed DAM-ROC framework could increase robustness against multiple evasion attacks without compromising on performance when compared against two existing frameworks, SLEIPNIR and KBL. GRAMS retrained Bayesian models have demonstrated consistent performance against all considered attacks which is crucial for real world scenarios since it cannot be predicted in advance which attack will be deployed.