張真臉：真活人臉驗證用於嵌入式系統

Abstract

在臉部識別的任務中，除了真實的人臉，可能會存在使用偽裝人臉的攻擊者，偽裝人臉的偵測任務就是為了偵測出使用偽裝者的資料，臉部偽裝者可能會有各種不同的攻擊手段，包含利用印刷照片、手機螢幕、面具，以及播放中的影片。要如何因應這些變化多端的攻擊手段是臉部偽裝偵測所需要面臨的課題。另外，不光是偽裝攻擊手段，環境的影響也是一個相當重要的因素，複雜或是光線變化劇烈的背景可能影響算法的判斷結果。因此，不因為環境而影響判斷才會是一個有足夠韌性的算法。

這篇論文對於偽裝人臉的偵測任務提出一個解決辦法。相對於傳統方法，深度學習的結果相對更好，並且在臉部偽裝偵測任務中，精確度是一項十分重要的衡量標準，因此這篇論文選擇使用深度學習技巧來提高對於臉部真偽分辨結果的可信度。關於臉部偽裝偵測的應用較多在於嵌入式系統上，像是手機、可視門鈴，因此在關注提高精確度之外，系統的運算能力以及算法所占的記憶體空間也視需要考慮的，這些也關乎嵌入式裝置的成本以及耗電量。

In face recognition, besides real human faces, attackers may use disguised faces. Face anti-spoofing detection intends to identify such data. Facial attackers may use various methods, including printed photographs, phone screens, masks, or even videos. How to address these ever-changing attack methods is a challenge faced by face recognition detection. Additionally, the environment is also a significant factor, as complex or rapidly changing lighting conditions may affect the algorithm decision-making. Therefore, a robust algorithm should not be affected by the environment during decision-making.

We propose ChangFAS for face anti-spoofing detection. Compared with traditional methods, deep-learning produces relatively better results. Accuracy is a crucial measure in face recognition detection, and therefore, We choose deep-learning techniques to improve the credibility of facial authenticity discrimination results.

Face anti-spoofing detection is mostly applied to embedded systems such as mobile phones and video doorbells. Therefore, besides focusing on improving accuracy, the system's computing power and the memory space occupied by the algorithm should also be considered, as they are related to the cost and power consumption of the embedded device.