基於單邊元三元組損失函數的多任務架構應用於泛化人臉防偽辨識

由於偽冒攻擊的變化日漸增加，模型的泛化對於人臉防偽辨識成為一項不可或缺的挑戰，然而先前許多提出的方法往往無法在泛化上表現得很好，本論文基於兩個角度來提升人臉防偽辨識的泛化能力，首先，在網路中使用人臉解析的資訊，讓網路能專注於臉部區域以及理解不同臉部區域的分布；第二點，使用單邊元三元組損失函數與元學習的過程進行合作。本論文提出一個新穎的多任務架構應用於泛化人臉防偽辨識，方法中包括三個任務: 深度預估、人臉解析、欺騙分類。藉由在人臉解析以及深度預測的任務中做像素級的監督，讓學到的特徵有正規化的效果，能夠更準確的區分出遭受攻擊的臉。另外，我們提出的單邊元三元組損失函數使用兩階段加大的邊界值，與元學習中模擬域名轉移的過程互相結合，增進模型泛化的能力。本論文提出的架構包括一個特徵提取器、一個深度預估器、一個基於U-net 的人臉解析器、以及一個元學習器負責元學習和分類器。而本文提出的基於U-net 的人臉解析器包含一個用來預估臉部語義照片的U-net，和一個基於注意力模型的連接，用來整合不同維度中的臉部語義資訊。本文中使用四個公開資料集來做測試泛化能力的實驗，證實我們提出的多任務架構以及訓練方法可以比先前其他方法在泛化的能力上表現得更好，面對沒看過的資料也有相當優越的結果，在一些人臉防偽辨識的域泛化基準實驗中，我們的方法相較於所參考的方法，AUC 進步超過6%，而比起過去的方法，HTER 也有相當的進步。

關鍵字

電腦視覺；深度學習；人臉防偽辨識；多任務；元學習；域名泛化

並列摘要

Due to increasing variations of presentation attacks, model generalization becomes an essential challenge for face anti-spoofing. Many previous works could not perform well in generalization. This paper improves the generalization ability of face anti-spoofing with two aspects. First, employing the face parsing information encourages the network to focus on face regions and realizes distributions between different face parts. Second, one-side triplet loss is adopted into the network to cooperate with the meta learning process. This paper proposes a novel multi-task face anti-spoofing framework that contains three tasks: depth estimation, face parsing, and live/spoof classification. With the pixel-wise supervision from the face parsing and depth estimation tasks, the regularized features can better distinguish spoof faces. While simulating domain shift with meta learning techniques, the proposed one-side triplet loss can further improve the generalization capability by a two-stage margin setting. Our framework consists of a feature extractor, a depth estimator, a U-net based face parsing module, and a meta learner for conducting meta learning and classification. The proposed U-net based face parsing module contains a U-net for predicting semantic face image and an attention-based skip connection for aggregating the semantic information of different channels. Extensive experiments on four public datasets demonstrate that the proposed framework and training strategies are more effective than previous works for model generalization to unseen domains. The AUCs are improved by over 6% compared to the baseline for some experiments on domain generalization benchmark for face anti-spoofing, and the HTER is also significantly improved over the previous methods.