Improved Detection of P2P Botnets through Network Behavior Analysis

Botnets are becoming powerful threats on the Internet because they launch targeted attacks towards organizations and the individuals. P2P botnets are resilient and more difficult to detect due to their nature of using different distributed approaches and encryption techniques. Classification based techniques proposed in the literature to detect P2P botnets, report high overall accuracy of the classifier but fail to recognize individual classes at the similar rates. Identification of non-bot traffic is equally important as that of bot classes for the reliability of the classifier. This paper proposes a model to distinguish P2P botnet command and control network traffic from normal traffic at higher rate of both the classes using ensemble of decision trees classifier named Random Forests. Further to optimize the performance, this model also addresses the problem of imbalanced nature of dataset using techniques like downsampling and cost sensitive learning. Performance analysis has been done on the proposed model and evaluation results show that true positive rate for both botnet and legitimate classes are more than 0.99 whereas false positive rate is 0.008.