基於重建類真實頻率的人臉偽造偵測技術

Abstract

現今基於深度學習的人臉偽造偵測技術大多藉由捕獲由偽造所形成的偽影來辨別影像的真偽。然而這類提取特定特徵的方法僅會在特定資料集表現出優異的性能，卻缺乏泛化的能力，所以在可能會出現新的或未知特徵的現實場景，將使得這類方法不再實用。人臉偽造通常會在頻域留下無法忽略的痕跡。在本研究中，我們提出了一種全新的臉部偽造偵測框架，透過學習真實人臉中的共同特徵，並以輸入圖像和重建圖像之間的距離來決定圖像的真偽。我們的方法首先破壞可疑的頻率，並在被破壞的頻帶中的進行資訊重建，使得重建影像更加接近真實，以改善對未知變造的泛化性。此外，我們提出了一種損失函數，將重建損失、度量學習損失和分類損失等相結合，以提高模型區分真偽的能力。我們在幾個基準資料集上對提出的方法進行實驗，數據顯示與最先進方法相比具有競爭力的結果。

Current learning-based face forgery detection methods typically discriminate between real and manipulated images by capturing artifacts caused by manipulation. However, while these artifact-based methods often perform well within a specific database, they lack the ability to generalize, making them impractical in real-world scenarios where new or unseen artifacts may be present. Fortunately, facial manipulations often leave non-negligible traces in the frequency domain. In this thesis, we propose a novel deepfake detection framework that learns the common features of real faces and uses the distance between an input image and its reconstruction image to discriminate between real and fake images. Our framework initially destructs suspicious frequencies and then reconstructs the information within the affected band to enhance the realism of the reconstruction and improve the generalization performance against unknown manipulations. In addition, we propose a loss function that combines reconstruction loss, metric learning loss, and classification loss to improve the model's ability to separate real from fake images. The proposed method is evaluated on several popular benchmark datasets and demonstrates competitive experimental results compared to state-of-the-art methods.