A Defensive Strategy Against Android Adversarial Malware Attacks

Abstract

Due to the popularity of Android mobile devices over the past ten years, malicious Android applications have significantly increased. Systems utilizing machine learning techniques have been successfully applied for Android malware detection to counter the constantly changing Android malware threats. However, attackers have developed new strategies to circumvent these systems by using adversarial attacks. An attacker can carefully craft a malicious sample to deceive a classifier. Among the evasion attacks, there is the more potent one, which is based on solid optimization constraints: the Carlini-Wagner attack. Carlini-Wagner is an attack that uses margin loss, which is more efficient than cross-entropy loss. We propose a model based on the Wasserstein Generative Adversarial Network to prevent adversarial attacks in an Android field in a white box scenario. Experimental results show that our method can effectively prevent this type of attack.

Keywords

• Adversarial attack
• Carlini-Wagner attack
• generative adversarial network
• android adversarial malware