Adversarial Deep Ensemble: Evasion Attacks and Defenses for Malware Detection

摘要: Malware remains a big threat to cyber security, calling for machine learning based malware detection. While promising, such detectors are known be vulnerable evasion attacks. Ensemble typically facilitates countermeasures, while attackers can leverage this technique improve attack effectiveness as well. This motivates us investigate which kind of robustness the ensemble defense or achieve, particularly when they combat with each other. We thus propose new approach, named mixture attacks, by rendering capable multiple generative methods and manipulation sets, perturb example without ruining its malicious functionality. naturally leads instantiation adversarial training, is further geared enhancing deep neural networks. evaluate defenses using Android against 26 different attacks upon two practical datasets. Experimental results show that training significantly enhances networks wide range promote base classifiers robust enough, yet evade enhanced effectively, even notably downgrading VirusTotal service.