Clustering the Unknown - The Youtube Case

Abstract

Recent stringent end-user security and privacy requirements caused the dramatic rise of encrypted video streams in which YouTube encrypted traffic is one of the most prevalent. Regardless of their encrypted nature, metadata derived from such traffic flows can be utilized to identify the title of a video, thus enabling the classification of video streams into a single video title using a given video title set. Nonetheless, scenarios where no video title set is present and a supervised approach is not feasible, are both frequent and challenging. In this paper we go beyond previous studies and demonstrate the feasibility of clustering unknown video streams into subgroups although no information is available about the title name. We address this problem by exploring Natural Language Processing (NLP) formulations and Word2vec techniques to compose a novel statistical feature in order to further cluster unknown video streams. Through our experimental results over real datasets we demonstrate that our methodology is capable to cluster 72 video titles out of 100 video titles from a dataset of 10,000 video streams. Thus, we argue that the proposed methodology could sufficiently contribute to the newly rising and demanding domain of encrypted Internet traffic classification.

Encrypted video traffic clustering demystified

“…1 exhibits a high-level view of the methodology as conducted in this study. As depicted, firstly a sanitisation module is triggered where re-transmissions and audio packets are eliminated from the video stream in a similar fashion as in the previous work in [7]. Moreover, Bits per Peak (BPPs) are computed for each encrypted video stream via re-assembling network flows by using the 5tuple packet-level information 4 between two Off periods within the observed peaks [7].…”

“…In order to confront the challenging task of clustering encrypted video with no a priori knowledge of video titles, the preliminary work in [7] exploited the usefulness of Natural Language Processing (NLP). The proposed solution was composed of a single clustering method relying on a language derived by meta-statistics of network traffic features per encrypted video stream.…”

“…Therefore, this work goes beyond previously introduced pieces of work and extends the preliminary findings 1. COBWEBS Technologies on EU GDPR policies: https://www.cobwebs.com/crime-investigations-eu/ in [7], by introducing a seminal synergy of NLP and Deep Auto Encoder-based feature embedding algorithms with refined clustering schemes such as grouping viewing profiles with unknown video streams. The empirical observation of the distribution of statistical features is neglected and an automated feature selection method with three well-known NLP techniques: Bag of Words, Auto-Encoder and a sample embedding using different affinity function is employed.…”

Explainable YouTube Video Identification Using Sufficient Input Subsets

“…However, numerous studies [3], [6] show that the combination of DASH and VBR is prone to leak information about video content. Patterns about particular videos that help in identifying those videos are learned by deep learning algorithms demonstrated in [6], [7], [8], and [9]. However, explaining the judgement of these models -models' rationale -for human understanding has remained a challenge for researchers.…”

Fingerprinting Technique for YouTube Videos Identification in Network Traffic

“…However, these segments are delivered to the client's device in a specific method, leaving a delivery pattern in the network traffic. This pattern can be used to identify the video in the network traffic [16], [17]. Many studies have exploited the DASH streaming pattern to identify the videos in the network traffic [18]- [21].…”