Individual Identification by Resting-State EEG Using Common Dictionary Learning

Abstract

Recently, a number of biometric methods to identify individuals based on personal characteristics, such as fingerprints and irises, have been developed. Individual identification based on electroencephalography (EEG) measurements is one of the safe identification techniques to prevent spoofing. In this study, we propose to employ common dictionary learning, which was formerly presented by Morioka et al. [1] aiming at performing subject-transfer decoding, for extracting features for EEG-based individual identification. Using the proposed method, though a classifier was trained based on the EEG signals during the selective spatial attention task, we found each test subject was almost perfectly identified out of 40 based on its resting-state EEG signals.

This study is originally presented in this manuscript. We previously submitted an abstract to a domestic technical meeting, but withdrew the submission so that we did not present the study at that meeting. The content page of the final meeting proceedings is attached below: http://www.ieice.org/ken/index/ieice-techrep-116-521-e.html.

Appendix: MDS of Spatial transforms \(Z_{ij}\)

Figure 3 displays multi-dimensional scaling (MDS) of $Z_{ij}$ over all subjects (40) and sessions ($1+8$) at the frequency band of 11.3 Hz. MDS presents the relative position in a low-dimensional space based on the distance information among data points (here, $Z_{ij}$), so that the more similar points are, the closer they are placed. We adopted the Frobenius norm between spatial transforms as distance information. Points are connected by colored lines if they are of a single subject. We can see that the points from a single subject (but different sessions) likely constitute a cluster, so that the distance between sessions of the same subject is shorter than that between different subjects regardless of session types. Considering the 2D-plot in Fig. 3 has some overlap due to the short of visualization dimensionality, these cluster structures are encouraging to make them use for subject identification. This result also shows the relevance of assumption (A3) that spatial transforms are consistent over different sessions of the same subject.

Fig. 3.

Multi-dimensional scaling of spatial transforms $Z_{ij}$ (11.3 Hz). Each point corresponds to a single session and points connected by colored lines are the spatial transforms of a single subject. Black and white points mean $Z_{ij}$ of resting and task sessions, respectively. (Color figure online)