Classifying network flows by their application type is the backbone of many crucial network monitoring and controlling tasks, including billing, quality of service, security and trend analyzers. The classical "port-based" and "payload-based" approaches to traffic classification have several shortcomings. These limitations have motivated the study of classification techniques that build on the foundations of learning theory and statistics. The current paper presents a new statistical classifier that allows real time classification of encrypted data. Our method is based on a hybrid combination of the k-means and knearest neighbor (or k-NN) geometrical classifiers. The proposed classifier is both fast and accurate, as implied by our feasibility tests, which included implementing and intergrading statistical classification into a realtime embedded environment. The experimental results indicate that our classifier is extremely robust to encryption.