使用視覺變換器於跨領域之人臉活體驗證

Abstract

隨著人臉辨識被大量地應用在生活中，與人臉活體驗證相關的研究也越發蓬勃，現行有許多方法能夠在特定的資料及上有很好的效果；然而在實際的應用上，人臉活體驗證模型應該要能夠被泛化到各種未知的領域 (domain)。近期 Vision Transformer 模型展現其在學習「具良好辨別性的特徵」的強大以及在許多視覺任務上有很好的表現，我們將它使用在跨領域的人臉活體驗證上。在這篇論文，我們首先簡述目前在人臉活體驗證上的研究方向，接著介紹我們提出的方法：Domain-invariant Vision Transformer (DiVT)。我們利用了兩項損失函數讓模型能夠有更好的泛化能力：為了讓模型能得知有多種方式被用於偽造影像，我們將真實資料以及使用不同攻擊手段的偽造資料各自集成群體，並將這些群體分離。其次，我們要求模型學習一個能夠適用於各個領域的特徵，並讓真實資料在特徵空間中能夠更集中，以幫助模型能更容易的區分出真實影像。實驗結果顯示，我們的方法達成了目前在跨領域之人臉活體驗證的最佳表現，且相較於之前的模型更簡單卻更有效。

Face recognition becomes more and more prevalent these days, which leads to numerous studies in face anti-spoofing (FAS). Existing FAS models have achieved high performance on specific datasets. However, for the application of real-world systems, the FAS model should generalize to the data from unknown domains rather than only achieve good results on a single baseline. As vision transformer models have demonstrated astonishing performance and strong capability in learning discriminative representations, we investigate applying transformers to distinguish the face presentation attacks over unknown domains. In this thesis, we first give a brief review of the research directions in FAS studies, then we propose the Domain-invariant Vision Transformer (DiVT) for FAS, which adopts two losses to improve the generalizability of the model. First, to make the model leverage the knowledge about the spoof data consisting of different types of attack, a separation loss is utilized to unify the real faces and the same types of attack, respectively. Then these groups are separated in the embedding space. Second, a concentration loss is employed to learn a domain-invariant representation that centralizes the features of real face data regardless of their domains, which helps the model to distinguish the real faces more easily. The experimental results show that our approach achieves state-of-the-art performance on the protocols of domain-generalized FAS tasks. Compared to previous domain generalization FAS models, our method is simpler but more effective.