利用視覺誘發腦電波之身份辨識

Abstract

近年來利用生物特徵的方式來進行身份辨識越來越普遍，其原因是由於生物特徵具有難以遭到破解或竊取的優點。然而，隨著科技的進步目前的生物特徵(例如：指紋、虹膜等)已有被複製的風險。由於腦電波具有個體間的差異，因此在本研究中我們利用視覺刺激誘發的腦電波為分析訊號來發展身份辨識系統，實驗於安靜無干擾的房間進行 ，讓受測者接受事件相關的視覺刺激(oddball paradigm)，利用刺激材料出現頻率的不同誘發出腦波的事件相關電位。辨識的步驟主要分為類別與確認兩大部分，並利用支援向量機作為分類器。類別的部分，原始訊號經過特徵擷取後藉由一個多種類的分類器會得到一個一對多的分類結果；而接著在確認部分，由類別步驟所得到的最佳分類結果經由此部分二元的分類器進行確認。特徵擷取方面，包含了降維，時域以及頻域的分析方法，能將具有代表性的資訊保留。此外，我們嘗試利用重複確認步驟的二元分類器將前一步驟(類別)分類錯誤的資料進行修正，修正的準則是依照支援向量機中的信賴評估為指標。 我們利用18位受測者的辨識結果得到97.25%的準確率，並且再經由確認的步驟能達到98.89%的正確接受率，這樣的結果顯示腦電波訊號具有的個體差異性足夠用於進行身份辨識且利用類別和確認兩部分的結合能達到一個好的準確率，且辨識的可信度提升。而更深入的討論訊號間的差異，我們發現不同受測者的訊號相關性低於同一受測者不同天的受測資料，這個發現符合腦電波具有低個體內差異性以及高個體間差異性，且隨著時間的變化同一人的訊號是恆定的。相關性的高低也解釋了某些受測者容易被錯誤分類的情況，也就是他們和其他人的訊號具有高度的相關性。總結我們系統所得出結果顯示，結合未來硬體發展更趨成熟腦波能成為一個新的生物特徵以發展成一套更安全的辨識系統。

The biometrics contains emerging methods for human identification. As advances in technology, conventional techniques using fingerprint or iris have the risk of being duplicated. In this work we utilize the inter-subject differences in the electroencephalographic (EEG) signals evoked by visual stimuli for person identification. The identification procedure is divided into classification and verification phases. For our classification system, it is based on the supervised classification method with support vector machine. During the classification phase, we extract the representative information from the EEG signals of each subject and construct a multi-class classifier. The best-matching candidate is further confirmed in the verification phase by using a binary classifier. The methods of feature extraction include dimension reduction and time-frequency analysis. Moreover, we try to correct those misclassified data through the iterative verification that depends on the confidence values of SVM classifier, which is a confidence level of classification. According to our experiments in which 18 subjects were recruited, the proposed method can achieve 97.25% identification rate. The results revealed that EEG data with individual differences can reach a high accuracy in person identification. Combining classification with verification, the reliability of the system can be increased. The correlation values of EEG signals between different subjects is lower than those of EEG signals acquired at different days for the same subject. This finding suggests that the characteristics of EEG has low intra-subject variability but high inter-subject variability and it is stable over time. The correlation values may also explain why some subjects apt to be misclassified when they have high correlation values to others. Our experimental results demonstrated that the proposed methods have great potentials for identifying individuals in daily life applications.