A Bucket-Based Data Pre-Processing Method for Encrypted Video Detection

As the number of video streaming platforms is growing, the risk factor associated with illegal and inappropriate content streaming is increasing exponentially. Therefore, mon- itoring such content is essential. Many researches have been conducted on classifying encrypted videos. However, most existing techniques only pass raw traffic data into clas- sification models, which is an ineffective way of training a model. This research proposes a bucket-based data pre-processing technique for a video identification in network traffic. The bucketed traffic is then incorporated with a fine-tuned word2vec-based neural net- work to produce an effective encrypted video classifier. Experiments are carried out with different numbers and sizes of buckets to determine the best configuration. Furthermore, previous research has overlooked the phenomenon of concept drift, which reduces the effec- tiveness of a model. This paper also compares the severity of concept drift on the proposed and previous technique. The results indicate that the model can predict new samples of videos with an overall accuracy of 81% even after 20 days of training.