Automated poisoning attacks and defenses in malware detection systems: An adversarial machine learning approach

摘要: Abstract The evolution of mobile malware poses a serious threat to smartphone security. Today, sophisticated attackers can adapt by maximally sabotaging machine-learning classifiers via polluting training data, rendering most recent machine learning-based detection tools (such as D rebin , roid APIM iner and M ) ineffective. In this paper, we explore the feasibility constructing crafted samples; examine how be misled under three different models; then conclude that injecting carefully data into significantly reduce accuracy. To tackle problem, propose K uafu et two-phase learning enhancing approach learns adversarial detection. includes an offline phase selects extracts features from set, online utilizes classifier trained first phase. further address environment, these two phases are intertwined through self-adaptive scheme, wherein automated camouflage detector is introduced filter suspicious false negatives feed them back We finally show boost accuracy at least 15%. Experiments on more than 250,000 applications demonstrate scalable highly effective standalone system.