運用腦電波之身分辨識系統及其長時調變機制

Abstract

生物特徵辨識因其廣泛性、可攜性與不易被複製等特質，而逐漸成為身分辨識的重要方法。運用腦電波進行生物特徵辨識是近年興起的研究主題，過去相關研究指出腦電波帶有個體獨特性的資訊且難以被竊取，故在高安全需求的系統上有相當高的應用潛力。然而腦電波辨識在準確度與穩定性上仍有改善空間。 本研究針對腦電波身分辨識系統提出了兩階段式身分辨識與長時調變機制，來分別改善準確率與穩定度。前者可在維持正確接受率的情況下，進一步降低錯誤接受率；後者則可因應腦電波特徵隨時間之變化來調變系統。兩階段式系統中包含了分類部分以及驗證部分。腦電波訊號經過自身回歸模型（autoregressive model, AR）與頻帶特徵的運算後，再透過多分類組別之支持向量機（support vector machine, SVM）來進行分類。而驗證部分是由候補者篩選(candidate selection)、資料重新表述(re-representation)和身分驗證三步驟所構成。我們首先運用線性區分分析（linear discriminant analysis, LDA）進行資料的重新表述，再由兩分類群組之支持向量機與最近相鄰分類器驗證分類結果。 另一方面，長時調變機制則分為漸進式學習（incremental learning）與在分類前先行調整資料分佈的適應性系統(adaptive system)。在適應性系統裡，我們探討了主成分分析與領域調整（domain adapatation）兩種方法。該系統是根據新資料來調整訓練資料的分佈，再利用調整過的內容重建分類器並進行身分辨識。 本研究共招募二十三位受試者，並在其自發性手指抬動的情況下收取90筆無眼動干擾之腦電波訊號。根據系統評估的實驗結果，兩階段式身分辨識能在兼顧正確接受率的情況下，有效地將錯誤接受率由9.5%降低至5.4%（正確接受率的改變幅度為90.5%略降為87.7%）。在穩定性的研究上，頻譜特徵是所有我們使用的特徵值中最不易隨時間改變的。若採用該特徵，主成分調整式身分辨識系統可增加5.2%的辨識正確率；漸進式學習系統更可將正確辨識率自12.2%提升至58.9%。

Biometrics has been viewed as an alternative to conventional person identification methods due to its universality, portability, and resistance in duplication. Previous study indicated that electroencephalography (EEG) carries discriminative information for distinguishing individuals and has great potential to meet the requirements of high security level. Therefore, it is essential to improve the reliability and stability of the EEG-based biometrics system for the promotion of its applicability. In this study, we first developed a two-stage EEG-based person identification system to improve the reliability by lowering the false acceptance rate while maintaining the true acceptance rate. The second purpose was to improve the stability of the EEG-based identification system using longitudinal adaptation. In addition to the classification stage, the proposed two-stage identification system re-examines the classification results in another verification stage, which consisted of candidate selection, re-representation, and verification steps. The classifier applied in the classification step was a multi-class support vector machine (SVM), while a binary SVM or kNN classifier was utilized in the verification step. Additionally, the linear discriminant analysis (LDA) was utilized to re-represent the training data before the verification step. On the other hand, the longitudinal adaptation was accomplished by either incremental learning or an adaptive system. In the adaptive system, we applied the principal component analysis (PCA) or domain adaptation (DA) to transform the data distribution toward the newly acquired data, and then we construct a new classifier with the adapted data. For each of the 23 participants, 90 EOG-free trials of the EEG recordings were acquired during the lifting of left index finger. The experimental results showed that our two-stage identification process reduced the false acceptance rate from 9.5% to 5.4% while maintaining the true acceptance rate (form 90.5% to 87.7%). The evaluation results showed that the power spectral density (PSD) was the most stable feature, and the PCA-based adaptive system can improve the true acceptance rate by 5.2% when using the PSD features. Furthermore, the identification system using incremental learning can even improve the true acceptance rate from 12.2% to 58.9%.