Evaluating Few Shot Learning With Uncertainty Quantification Under Encrypted Traffic Classification

Abstract

Computer networks frequently contain novel applications which a network administrator could benefit from understanding. Encryption and virtual-private-network (VPN) service providers are at odds with this. This thesis contains a benchmark and study of the Out-Of-Distribution (OOD) detection and classification of encrypted VPN traffic with few shot learning algorithms. The research determines the impactful hyperparameters of a few shot learning algorithms under the VPN setting. The work then moves to better understand the OOD detection performance, testing alternative few shot learners and finding potential trade offs between them. The research finds that a transductive, few shot learner has superior OOD detection to its inductive counterpart. However, transductive methods typically require more data to configure. Therefore, the research develops and tests a hybrid inductive-transductive approach, thus determining if a middle ground is possible without too many negative consequences.