Human activity detection and classification in network traces: Leverage the machine learning power

Details

Supervisors

Sadre, Ramin

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enThis thesis focuses on detecting and classifying human activity within connected environments by analyzing network traffic. The primary goal is to develop automated tools for law enforcement investigators to infer user behavior on surveilled networks. A processing pipeline was designed and implemented to predict activity types at any given moment, using only outgoing traffic from a household's internet gateway. This pipeline utilizes supervised learning models and explores two feature-extraction strategies: aggregating packets into fixed time intervals to capture temporal patterns, and grouping packets by flow identifiers to capture session-level behaviors. After training these models on a manually labeled dataset, both approaches demonstrated high accuracy in detecting human activity (over 95%). They also classified activities into predefined categories with accuracies of 87% for the time-window approach and 89% for the flow-based approach, although performance varied for underrepresented classes. These findings confirm that significant information about user behavior can be extracted from encrypted traffic, and that the novel time-window aggregation performs comparably to traditional flow-based methods. Our performance is in line with what is done in the field, confirming that our model is effective in our case study.