CGTN｜AI生成的深度伪造泛滥，如何识别？

How to spot AI-generated deepfakes

21-Mar from CGTN

AI fakery is quickly becoming one of the biggest problems confronting us online. Deceptive pictures, videos and audio are proliferating as a result of the rise and misuse of generative artificial intelligence tools. Video and image generators like DALL-E and OpenAI's Sora make it easy for people without any technical skills to create deepfakes – just type a request and the system spits it out. These fake images might seem harmless. But they can be used to carry out scams and identity theft or propaganda and election manipulation.
人工智能造假正迅速成为我们在网上面临的最大问题之一。由于生成人工智能工具的兴起和滥用，欺骗性的图片、视频和音频正在激增。像 DALL-E 和 OpenAI 的 Sora 这样的视频和图像生成器，让没有任何技术的人也能轻松制作深度伪造品--只需键入一个请求，系统就会自动生成。这些虚假图片看似无害。但它们可以被用来实施诈骗和身份盗窃，或进行宣传和操纵选举。

A lot of AI deepfake photos, especially of people, have an electronic sheen to them, Ajder said, a leading expert in generative AI. Check the consistency of shadows and lighting. Often the subject is in clear focus and appears convincingly lifelike but elements in the backdrop might not be so realistic or polished.
生成式人工智能的权威专家Ajder说：很多人工智能深度伪造的照片，尤其是人物照片，都带有电子光泽。检查阴影和照明的一致性。通常情况下，主体对焦清晰，看起来栩栩如生，但背景元素可能并不那么逼真或精致。

Face-swapping is one of the most common deepfake methods. Does the facial skin tone match the rest of the head or the body? Are the edges of the face sharp or blurry? Do their lip movements match the audio perfectly? Cybersecurity company Norton says algorithms might not be sophisticated enough yet to generate individual teeth, so a lack of outlines for individual teeth could be a clue. There are tools online that promise to sniff out fakes. Microsoft has developed an authenticator tool that can analyze photos or videos to give a confidence score on whether it's been manipulated.
换脸是最常见的深度伪造方法之一。面部的肤色与头部或身体的其他部位匹配吗？脸的边缘是尖锐的还是模糊的？他们的嘴唇动作与音频完美匹配吗？网络安全公司Norton表示，算法可能还不够复杂，无法生成单个牙齿，因此缺乏单个牙齿的轮廓可能是一个线索。有一些网上工具可以识别出假货。微软已经开发出一种验证工具，可以分析照片或视频，从而对照片或视频是否被篡改给出一个可信度评分。

However, AI models are being trained on internet data to produce increasingly higher-quality content with fewer flaws. That means there's no guarantee this advice will still be valid even a year from now.
然而，人工智能模型正在接受互联网数据的训练，以产生质量越来越高、缺陷越来越少的内容。也就是说，即使一年后，也不能保证这些建议仍然有效。

今日词汇
spot 发现，识别
deepfake 深度伪造
fakery 伪造，假货
confront 面临，面对
deceptive 迷惑的，欺骗性的
proliferate 增殖，激增
scam 欺诈，骗局
identity theft 身份盗用
propaganda 宣传，鼓吹
sheen 光泽，光彩
consistency 连贯性，一致性
lifelike 生动的，逼真的
Face-swapping 换脸
skin tone 肤色
blurry 模糊不清的
algorithm 算法，计算程序
sniff out 闻出，发觉
authenticator 认证者，验证器