數位影像編輯工具可輕易改變影像甚至視訊內容，並同時保持極高的畫面品質。深度偽造(DeepFake)的出現造成更大的影響，也因為各種惡意目的的操作下，這些對於內容的篡改使數位影像與視訊的真實性帶來威脅與挑戰。近年來有不少偵測畫面內容竄改的方法被提出，大多採用機器學習或深度學習相關技術。然而竄改方式的多樣與不斷演進變化，讓搜集所有型態的竄改資料以進行監督式訓練變得困難或不切實際，即便蒐集齊全也可能面臨資料集過於龐大而需要更多訓練資源等問題。 本研究從另一個角度出發，提出基於區塊相似性的深度學習辨識方法，透過評估區塊內容的一致性來判斷影像或視訊中的偽造或受影響區域。這種方法旨在避免蒐集各類型竄改資料進行訓練，我們選擇使用原始或未修改畫面區塊來實現相關的辨識與偵測。我們訓練一個卷積神經網路來提取影像區塊特徵，使用孿生(Siamese)網路進行區塊對之間的相似度比對，以確定畫面中可能被竄改的區域。對於影像竄改偵測，我們另引入分割網路以對竄改區域進行進一步精細處理。對於深偽視訊偵測，我們首先定位人臉區域，接著透過比對前後幀中的人臉區域相似度來判斷該視訊的真實性。我們在公開資料集上對所提出的方法進行測試和驗證，以證實所提出方法的可行性。這些資料集包含各種不同類型的影像與視訊，涵蓋多種內容竄改操作。與其他方法比較顯示所提出方案在準確性和穩定性的優越。 ;Digital image editing tools enable effortless manipulation of images and video content while maintaining high visual quality. However, the emergence of DeepFake has introduced significant challenges to the authenticity of digital media. Various methods for detecting such content manipulations have been proposed, primarily relying on machine learning or deep learning techniques. However, the constantly evolving nature of manipulation methods makes it impractical to collect all types of manipulated data for supervised training. Additionally, handling large datasets can be resource-intensive. In this study, we propose a deep learning-based method that utilizes block similarity to identify forged or manipulated regions within images or DeepFake videos by evaluating the consistency of block content. Our approach aims to avoid the need for collecting various types of manipulated data for training. Instead, we opt to use original or unmodified blocks for forgery detection. We train a convolutional neural network to extract features from image blocks and employ a Siamese network to compare block similarity. For image manipulation detection, we introduce a segmentation network to further refine the detection of manipulated regions. In the cases of DeepFake video detection, we first locate facial regions and then determine the video′s authenticity by comparing facial region similarity between consecutive frames. We conduct tests on publicly available datasets, encompassing images and videos with various content manipulation operations. The experimental results demonstrate superior accuracy and stability compared to other existing methods.
