Authors: Hemant Rathore, Adarsh Nandanwar, Sanjay K. Sahay, Mohit Sewak
DFRWS EU 2023
Today, android smartphones are being used by billions of users and thus have become a lucrative target of malware designers. Therefore being one step ahead in this zero-sum game of malware detection
between the anti-malware community and malware developers is more of a necessity than a desire. This work focuses on a proactive adversary-aware framework to develop adversarially superior android malware detection models. We first investigate the adversarial robustness of thirty-six distinct malware detection models constructed using two static features (permission and intent) and eighteen classification algorithms. We designed two Targeted Type-II Evasion Attacks (TRPO-MalEAttack and PPO-Mal-EAttack) based on reinforcement learning to exploit vulnerabilities in the above malware detection models. The attacks aim to add minimum perturbations in each malware application and convert it into an adversarial application that can fool the malware detection models. The TRPO-MalEAttack achieves an average fooling rate of 95.75% (with 2.02 mean perturbations), reducing the average accuracy from 86.01% to 49.11% in thirty-six malware detection models. On the other hand, The PPO-MalEAttack achieves a higher average fooling rate of 96.87% (with 2.08 mean perturbations), reducing the average accuracy from 86.01% to 48.65% in the same thirty-six detection models. We also develop a list of the TEN most vulnerable android permissions and intents that an adversary can use to generate more adversarial applications. Later, we propose a defense strategy (MalVPatch) to counter the adversarial attacks on malware detection models. The MalVPatch defense achieves higher detection accuracy along with a drastic improvement in the adversarial robustness of malware detection models. Finally, we conclude that investigating the adversarial robustness of models is necessary before their real-world deployment and helps achieve adversarial superiority in android malware detection.