What is Deepfake? 「深度偽造」是什麼？

【摘要2019.11.4.自由】◎劉宜庭

Deepfakes rely on a branch of AI called Generative Adversarial Networks (GANs). It requires two machine learning networks that teach each other with an ongoing feedback loop. The first one takes real content and alters it. Then, the second machine learning network, known as the discriminator, tests the authenticity of the changes.

「深度偽造」依賴人工智慧（AI）的分支「生成對抗網路」（GAN）。「生成對抗網路」責成兩個機器學習網路，透過不斷生成的回饋迴路互相學習。第一個「機器學習」網路變造真實的內容。然後，第二個機器學習網路「資料辨識器」，對這些更動進行鑑識。

GANs are still in the early stages, but people expect numerous potential commercial applications. For example, some can convert a single image into different poses. Others can suggest outfits similar to what a celebrity wears in a photo or turn a low-quality picture into a high-resolution snapshot.

「生成對抗網路」仍處於初步發展階段，但人們預期它在商業應用上擁有巨大潛力。例如，它可以將單張影像變換成不同的姿勢。它們可以找出名人在照片中所穿的類似服裝，或將低畫質的照片轉換為高解析度的快照。

But, outside of those helpful uses, deepfakes could have sinister purposes. Consider the blowback if a criminal creates a deepfake video of something that would hurt someone’s reputation — for instance, a deepfake video of a politician "admitting" to illegal activities, like accepting a bribe.

然而，除了正向的應用之外，也會有居心不良的「深度偽造」。倘若罪犯創造出一部妨害他人名聲的深度偽造影片，例如讓一名政治人物「承認」接受賄賂等違法行為的一部深偽影片，所造成的衝擊將難以衡量。

新聞辭典

deepfake：名詞，深度偽造、深偽。例句：In one of the best-known examples of Deepfake manipulation, BuzzFeed published a clip now widely known as "ObamaPeele."（「深度偽造」操弄的著名案例之一，是BuzzFeed發布的「歐巴馬皮爾」影音片段。）

machine learning：專有名詞，機器學習。例句：Machine learning is the science of getting computers to act without being explicitly programmed.（「機器學習」是一種讓電腦不需要明確編程即可運行的科學。）

GAN（Generative Adversarial Network）：專有名詞縮寫，生成對抗網路。例句：GANs were introduced in a paper by Ian Goodfellow and other researchers at the University of Montreal in 2014.（「生成對抗網路」在2014年的一篇論文中，首次由伊恩．古德菲洛等蒙特婁大學的研究人員提出。）