Evasion Attacks On Svm Classifier

Abstract

Support Vector Machine (SVM) is one of the main classification techniques used in many security-related applications like malware detection, spam filtering, etc. To incorporate SVM into real-world security applications they must be able to cope up with the attack patterns that will lead to misclassifications. In this system, the vulnerability of SVM to evasion attacks are measured. A simple but effective approach is presented that can be exploited to systematically assess the security of widely-used classification algorithms against evasion attacks. To identify the vulnerabilities some transformations are applied to the testing set of handwritten digit images. The obtained result is plotted as a confusion matrix that allows the visualization of the performance of the algorithm against evasion attack. The work demonstrates the correctness and performance of existing adversarial systems. This work also compares the performance level of feature descriptors like Speeded Up Robust Features (SURF) and Histogram of Oriented Gradients (HOG) and their level of vulnerability to the evasion attacks are also measured. It can be inferred from our system that, even though both HOG and SURF are vulnerable to evasion attacks, those images that are extracted using SURF is less vulnerable compared to those images extracted using HOG features.