Multi-Task Attention Network for Digital Context Classification from Internet Traffic
Wenbin Li, Doua Hugues ZEMI, Valentin REDON, Matthieu Liewig · 2022
Understanding multi-aspect of user's digital context over Internet is a challenging due to the increasing network data encryption, privacy protection and complexity of the context concept. Based on the emerging attention mechanism capturing models’ focus points of input data, we proposed in this paper the attention neural network for digital context classification from Internet traffic: firstly, we introduced a digital context model and an extended version of multi-label dataset of Internet traffic to support our classification work. In the following, single task models and a multi-task model were proposed, and a series of experiments on five classification tasks of digital context has been conducted to evaluate the performance of recurrent models, attention models and the multi-task model. Experiments’ results have shown that different attention mechanisms are able to improve the outcome of recurrent classification models for Internet traffic, while multi-task attention network is able to further improve the overall accuracy of all tasks in the model by mutualizing data and exploiting implicit relations among tasks.