Passive Monitoring of HTTPS Service Use

HTTPS is used today to secure the majority of web communications and so enhance user privacy. Therefore, traffic monitoring techniques must evolve to remain useful, especially to support security considerations, as for example detecting and filtering the forbidden uses of a web service. However, privacy should remain as intact as possible. This paper describes a new passive and transparent method to infer the use of a HTTPS service by extracting and interpreting only meaningful meta-data derived from the encrypted traffic without deeply profile individual users. We propose a model using the sizes of objects loaded in the HTTPS service as a signature, by leveraging kernel density estimation, supporting then a classification function. We assess this approach extensively on the Google Images Service although our approach remains valid for some other services. We succeed to achieve an accuracy of 99.18% when detecting particular keywords to be searched over a large dataset of 115,500 distinct keywords.