近年來,圖像生成以及圖像屬性改造的方法能產生以假亂真的結果,使其有可能被用來當作製造假新聞的工具。作為反擊,學術圈也投注更多的心力在自動偵測假圖片上面。本論文提出利用半監督異常偵測的方法來增進偵測未知造假方式的能力,此方法在特徵空間中縮小真實圖像和中心的距離並且放大偽造圖像和中心的距離。和一般二分類問題不同的是,此方法並沒有強制讓偽造的圖像聚成一類。我們相信這是有幫助的,因為他體現了不同造假方式的差異性。我們將此方法應用在偵測卷積神經網路產生的圖像和偽造人臉的資料集,實驗中發現共同使用此方法和標準二元分類用的交叉熵能增進模型的通用性,我們相信這個效能增進說明了此方法的可能性。
Recent photo-realistic image synthesis or attribute manipulation methods could be used as a tool to create fake news. As a fightback, researchers start to put more effort on automatically detecting fake images. This thesis proposes to leverage an objective in semi-supervised anomaly detection to increase the generalization ability on detecting images generated from unseen forgery methods. This objective tries to minimize the distances between real samples and the center and maximize the distances between fake samples and the center at the same time. Unlike standard binary classification, it doesn't assume that all the fake images should be near to each other in the feature space. We believe that this is helpful for generalization since it reflects the diversity of different fake methods. We examine this method on detecting CNN-generated images and fake faces. The improvement of the generalization ability can be found when incorporating with binary cross-entropy loss. The performance gains in experiments show the potential of this method.