Electroencephalography signal based subject identification using imagined speech

Biometric identification systems have become key components in data security andprotecting sensitive information. Biometric methods, such as fingerprint recognition, have replaced traditional authentication methods due to their high security and effi-ciency. However, challenges like the potential to forge have highlighted the need for the development of more robust methods. A new approach in this field is the use ofelectroencephalography signals for identity verification, which not only provides high security but can also enhance the safety of brain-computer interfaces. In this study, we introduce a cueless imagined speech paradigm based on natural word selection, where users select and imagine semantically meaningful words withoutreceiving any external visual or auditory cues. This method overcomes the limitations of previous approaches and provides more realistic conditions for data collection. Based onthis approach, a dataset comprising over 4,350 samples from 11 individuals (7 males and 4 females) across five sessions was gathered. These sessions, held with specified intervals,were conducted within a single day to investigate the impact of electroencephalography signal variations over time. This approach brings data collection conditions closer toreal-world applications. For data processing, an automated preprocessing framework was designed by com-bining methods from the literature. Additionally, two-stage and end-to-end classifica- tion frameworks were optimized and evaluated by combining the models in the litera-ture, including the feature fusion of the MOMENT foundation model with a Support Vector Machine classifier, as well as deep learning models like EEG Conformer andShallow ConvNet. A reliable validation approach was used to ensure valid evaluation and prevent data leakage. Unlike some previous studies that selected training and testsamples from the same session, this study ensured that the training and test samples were chosen from separate sessions with time intervals. Results showed that after hy-perparameter optimization and model comparison, the EliteVote model, an ensemble of the top three models in this study using a majority voting approach, achieved thehighest accuracy of 98.31. Furthermore, due to its greater robustness, it was selected as the final classifier for the identity verification system.