TriNet-MTL: A Multi-Branch Deep Learning Framework for Biometric Identification and Cognitive State Inference from...

Auditory-evoked EEG signals contain rich temporal and cognitive features that reflect both the identity of individuals and their neural response to external stimuli. Traditional unimodal approaches often fail to fully leverage this multidimensional information fully, limiting their effectiveness in real-world biometric and neurocognitive applications. This study aims to develop a unified deep learning model capable of jointly performing biometric identification, auditory stimulus language classification, and device modality recognition, thereby exploiting both physiological and cognitive dimensions of auditory-evoked EEG. We introduce TriNet-MTL (Triple-Task Neural Transformer for Multitask Learning), a multi-branch deep learning framework composed of a shared temporal encoder and a transformer-based sequence modeling unit, trained and validated on auditory-evoked EEG data from 20 human participants (16 males and 4 females). The architecture is designed to simultaneously learn task-specific features via three dedicated output heads, each addressing one of the following: user identity (biometric), stimulus language (native vs non-native), and stimulus delivery mode (in-ear vs bone conduction). The model is trained using a sliding window approach and optimized through joint cross-entropy loss across tasks. TriNet-MTL demonstrates robust performance across all three classification tasks, achieving high accuracy in biometric identification (>93%) and strong generalization in cognitive state inference. Multi-task training further improves representation learning, reducing inter-task interference while enhancing task synergy. The proposed TriNet-MTL framework effectively captures both user-specific and cognitively informative patterns from auditory-evoked EEG, establishing a promising direction for integrated EEG-based biometric authentication and cognitive state monitoring in real-world systems.