基於無監督深度特徵分群法的人臉影像防偽辨識模型

Abstract

隨著臉部辨識在各種授權系統的使用需求趨於增加，如何防止非法入侵者透過臉部影像造假技術而得到使用許可變成重要的研究議題。在過往研究者的努力之下，在同一臉部防偽資料集內的訓練及測試準確率，已經達到可觀的研究成果。但是，根據某一資料集所訓練的模型，再將之應用到其他資料集時，仍存在許多問題；也就是在跨資料集測試時，錯誤率會有巨大的增加。為了解決此問題，本論文提出了一種無監督及遷移學習的方法，以提高模型在跨資料集時的泛化能力。我們使用預訓練深度學習模型來抽取臉部真實及造假影像的高維影像特徵，並透過降維及無監督分群方法取得針對測試資料集的分類結果。此方法的特色在於，預訓練模型的訓練資料不只使用臉部防偽影像，也納入模糊偵測及自然影像資料集。在我們提出的模型架構中，甚至可以在不使用臉部防偽影像的情況下完成訓練。與現有的遷移學習中，大多仍會採用目標資料集的標籤來調整模型的方法相比，本方法更偏重於尋找足以多重跨資料集導向的模型。在現有主要的公開資料集實驗中，我們的分類模型的平均錯誤率，比起現有最好方法低了3%以上，我們相信無監督深度分群的模型對於克服此一問題有極大的研究潛力。

With the increasing requirements for face recognition in many authentication systems, how to prevent intruders from accessing the permission via Face Anti-Spoofing（FAS） techniques has become an important research area in biometrics. After the endeavors over the past few years, researchers around the world have achieved acceptable FAS detection accuracy in the same training and testing dataset. However, it is still problematic when the model trained on one dataset is tested on some other datasets. The detection error rate increases dramatically when this kind of cross-dataset evaluation arises. To address this issue, this thesis introduces the unique techniques of transfer learning and unsupervised learning to increase the generalization ability for cross-dataset evaluation. Specifically, we develop a pre-trained deep learning model to extract the high dimension features of the attack and bona fide images, and the extracted features are clustered into two subsets after the dimension is reduced. One particular characteristic of this strategy is that the dataset that being used to train the pre-trained model is not necessarily in the FAS domain, which makes our framework naturally cross-data oriented. This is quite different from other existing transfer learning methods, which mostly utilize the labeled data of the target domain to fine-tune the model parameters. Based on benchmark dataset experiments, our FAS classifier achieved lower average classification error rate (ACER) scores than state-of-the-art methods by 3%. We believe that the proposed semi-supervised learning model is of potential to overcome this challenging FAS task in biometrics.