Apprentissage profond pour l’authentification continue et la reconnaissance d’activités humaines à partir de données comportementales multimodales

Abstract

Modern smartphones are equipped with a variety of sensors, such as the accelerometer, gyroscope, and capacitive touchscreen, which offer a unique opportunity for continuous authentication and human activity recognition. This thesis proposes two deep learning and structured state-space model (SSM) based architectures : S4HI, for robust identification using inertial behavioral data, and MEDUSAA, a multimodal encoder-decoder that combines inertial sensors and tactile interactions. These approaches leverage the ability of SSMs to effectively capture long-term temporal dependencies while reducing complexity and inference time compared to Transformertype architectures. The models are evaluated on several public datasets (HMOG, OU-ISIR, WhuGait, PAMAP2, Capture-24) and demonstrate superior performance in terms of precision, scalability, and online deployment. The results confirm the potential of these methods for secure, low-energy embedded applications.

Date	28 Aug 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Waël Jaafar (Supervisor) & Rami Langar (Co-supervisor)