| 摘要: | 本論文提出了一種結合 Blended 和光流圖的深度偽造辨識模型，對於提升偽造手法
的識別能力提供了有效的改進。在深度學習快速發展的背景下，除了熟知的圖像識別
應用，推薦系統和醫療診斷等領域也深刻影響著我們的日常生活。然而，這些技術進
步的同時也伴隨著潛在的風險，例如可能威脅隱私和安全的深偽技術。
深偽技術因其生成虛假圖像和合成真實影片的能力而成為一個日益嚴峻的問題。
隨著技術不斷演進，偽造方法也在推陳出新，導致訓練深偽檢測模型面臨諸多挑戰，
特別是需要大量資料來支持模型訓練，而新型偽造方法的數據獲取往往困難重重。
深偽技術的快速發展引發了關於交叉偽造（cross-manipulation）問題的重要議題。
交叉偽造指的是模型在經過針對特定偽造手法的訓練後，必須能夠對抗不同的未知偽
造手法，這對模型的泛化能力提出了更高要求，進一步增加了辨識真偽的挑戰，而又
衍伸出跨資料集（cross-dataset）的問題，除了要能夠識別不同手法也需要處理因不同
資料集產生分布上的差異導致深偽檢測準確度下降的問題。
為了解決跨資料集（cross-dataset）問題，本研究提出了一種結合動態與靜態特
徵的偽造檢測方法。該模型不僅在 cross-dataset 測試中顯著提升了檢測效能，還在
cross-manipulation 場景中保持了穩定且高效的檢測能力。
實驗結果證明了動態與靜態特徵結合的優勢：動態特徵能捕捉連續幀間的微妙變
化，特別是偽造技術引入的時間域異常；靜態特徵則有效提取單幀圖像的局部細節與
紋理信息。通過整合這兩類特徵，模型能更加準確地識別真偽，在交叉偽造場景中展
現出卓越的分類能力與更高的泛化性。;This paper proposes a deepfake detection model that combines blended and optical flow
features, offering effective improvements in the ability to identify various forgery methods.
In the context of rapid advancements in deep learning, its applications extend beyond image
recognition to domains such as recommendation systems and medical diagnostics, profoundly
impacting daily life. However, alongside these advancements come potential risks, such as the
misuse of technologies that threaten privacy and security, exemplified by deepfake technology.
Deepfake technology has become an increasingly pressing issue due to its ability to gen-
erate fake images and synthesize realistic videos. As the technology evolves, forgery methods
continue to emerge, posing challenges to training deepfake detection models. These challenges
include the need for large datasets to support model training, while acquiring data for novel
forgery methods remains a significant obstacle.
The rapid development of deepfake technology has raised critical concerns about the cross-
manipulation problem. Cross-manipulation refers to the requirement for a model trained to
recognize specific types of forgeries to generalize effectively against unseen forgery methods.
This necessitates stronger generalization capabilities, further complicating the task of identifying
authenticity. Additionally, this issue extends to the cross-dataset problem, where the model must
not only identify various forgery techniques but also address distributional differences between
datasets that lead to a decline in detection accuracy.
To address the cross-dataset problem, this study proposes a forgery detection method that
integrates dynamic and static features. The model demonstrates significant improvements in
detection performance during cross-dataset testing and maintains stable and efficient detection
capabilities in cross-manipulation scenarios.
Experimental results highlight the advantages of combining dynamic and static features.
ii
Dynamic features capture subtle variations between consecutive frames, particularly temporal
anomalies introduced by forgery techniques. Meanwhile, static features effectively extract local
details and texture information from individual frames. By integrating these two types of fea-
tures, the model achieves more accurate forgery detection, demonstrating robust classification
performance and superior generalization in cross-manipulation scenarios.
In summary, this study provides an innovative and robust solution to deepfake detection,
addressing both cross-dataset and cross-manipulation challenges, and laying a solid foundation
for future research in this critical area. |
