부산대학교 도서관

Malware identification is a difficult task that has been recently approached by training classifiers through machine learning. We present here a low complexity semi-supervised dictionary learning framework that begins with training an initial dictionary on a small labeled data set, and then continues with online learning on incoming unlabeled data, making use of every sample that it is exposed to, with the scope of adapting to new and unknown malware types. Our main contribution is a new online algorithm that makes use of regularization techniques that balance the capability of the dictionary to express both fresh and well established patterns.