Fine-grained TLS services classification with reject option

The recent success and proliferation of machine learning deep have provided powerful tools, which are also utilized for encrypted traffic analysis, classification, threat detection in computer networks. These methods, neural networks particular, often complex require a huge corpus training data. Therefore, this paper focuses on collecting large up-to-date dataset with almost 200 fine-grained service labels 140 million network flows extended packet-level metadata. number is three orders magnitude higher than other existing public labeled datasets traffic. labels, important to make the problem hard realistic, four times most class labels. published intended as benchmark identifying services Service identification can be further task "rejecting" unknown services, i.e., not seen during phase. Neural offer superior performance tackling more challenging problem. To showcase dataset's usefulness, we implemented multi-modal architecture, state-of-the-art approach, achieved 97.04% classification accuracy detected 91.94% 5% false positive rate.