Malware Classification and Defence Against Adversarial Attacks

In today’s world, new kinds of malicious codes are being created every day. This malware have the potential to compromise our systems and cause tremendous loss, may it be hardware based, or data oriented. We felt that there was a need to design a robust mechanism that can defend our systems against such malicious attacks. Traditional methods can’t cope up with new kinds of malware. Machine learning would not only help in detecting the known kinds of malware, but also in identifying new kinds of malware. So, we implemented various machine learning techniques to make accurate detection of malware. We proposed the use of a hybrid classifier for further improving the accuracy of malware classification in comparison to the existing machine learning techniques. Although machine learning does provide a solution, it is still vulnerable to adversarial attacks. Hence, we created adversarial examples and analyzed their impact on machine learning classifiers. Then, as a defence method, we implemented adversarial training where the machine learning classifiers were trained on the adversarial examples along with the original samples. Adversarial training enabled the classifiers to become robust against potential adversarial attacks.