Adversarial machine learning in malware detection

U današnje vrijeme sve je popularnija zaštita od malignih programa pomoću algoritama strojnog učenja koji trenutno rade prilično dobro. Međutim, kako to obično biva, kad god dizajneri uspiju pokrpati neki postojeći sigurnosni problem, napadači se odmah daju u posao za traženjem novih sigurnosnih propusta. Istovremeno, dizajneri pokušavaju unaprijed dokučiti koji bi mogao biti propust te i sami testiraju vlastite klasifikatore te smišljaju načine na koje bi ih srušili. Iz toga se razvilo relativno mlado područje računarstva, mješavina strojnog učenja i računalne sigurnosti – suparničko strojno učenje. Ono se bavi traženjem sigurnosnih propusta u algoritmima za zaštitu od neželjenih programa koji se baziraju na strojnom učenju, te popravljanjem istih, nalaženjem načina da se ti isti napadi spriječe, detektiraju i ne ponavljaju. Nakon razmatranja i provođenja klasične klasifikacije skupa benignih i malignih aplikacija, konstruiran je algoritam na bazi suparničkog strojnog učenja koji je modifikacijom malignih aplikacija pokušavao progurati iste kroz klasifikator bez da budu detektirani. Provedeni su eksperimenti na dva seta podataka koji su pokazali vrlo zanimljive rezultate i brojne nesigurnosti u postojećim jednostavnim algoritmima za klasifikaciju malignih i benignih programa – klasifikatorima LinearSVC i MultinomialNaiveBayes. Ostavlja se prostor za daljnje rasprave, razrade, poboljšanja postojećeg napadačkog algoritma te algoritma za obranu od provedenog napada.Nowadays, protection against malicious software by using machine learning is more and more popular, especially because of their exceptional efficacy in fighting malware. However, as usual, whenever designers succeed in coping with an existing security problem, the attackers immediately get busy with searching for new security flaws. At the same time, designers are trying to figure out in advance what could be a potential failure or a flaw to their system, and test their own classifiers themselves, figuring out how the potential attackers would be able to crash them down. From the arms race that has been just described, a relatively small area of computing has developed, a mixture of machine learning and computer security - adversarial machine learning. It consists of the search for security vulnerabilities in malware protection algorithms based on machine learning, and of the process and research of the ways to repair these algorithms in order for them to be able to detect, prevent and fight those same attacks. After contemplating and implementing a basic classification of a set of benign and malignant applications, an adversarial algorithm was constructed, which tried to push the malware through the classifier without it being detected. The algorithm did so by modificating the feature vectors of given malware. Experiments were conducted on two sets of data that showed very interesting results and numerous uncertainties in the existing simple algorithms for classifying malicious and benign programs - LinearSVC and MultinomialNaiveBayes classifiers. There is room left for further discussion, elaboration, improvement of the existing attack algorithm and for creating the algorithm for defense against the attack