Enhancing Adversarial Examples for Evading Malware Detection Systems: A Memetic Algorithm Approach

Malware detection using Machine Learning techniques has gained popularity due to their high accuracy. However, ML models are susceptible to Adversarial Examples, specifically crafted samples intended to deceive the detectors. This paper presents a novel method for generating evasive AEs by augmenting existing malware with a new section at the end of the PE file, populated with binary data using memetic algorithms. Our method hybridizes global search and local search techniques to achieve optimized results. The Malconv Model, a well-known state-of-the-art deep learning model designed explicitly for detecting malicious PE files, was used to assess the evasion rates. Out of 100 tested samples, 98 successfully evaded the MalConv model. Additionally, we investigated the simultaneous evasion of multiple detectors, observing evasion rates of 35% and 44% against KNN and Decision Tree machine learning detectors, respectively. Furthermore, evasion rates of 26% and 10% were achieved against Kaspersky and ESET commercial detectors. In order to prove the efficiency of our memetic algorithm in generating evasive adversarial examples, we compared it to the most used evolutionary-based attack: the genetic algorithm. Our method demonstrated significantly superior performance while utilizing fewer generations and a smaller population size.