研究：Fiverr 上约 600 个服务在售 AI 生成 NSFW 深度伪造内容
AI-Generated NSFW Services Are Entering Mainstream Marketplaces | Mohamed Moustafa Dawoud
一项针对 Fiverr 的测量研究（USEC 2026）发现，该平台存在约 600 个宣传 AI 生成 NSFW 内容的服务，其中 82.8% 提供换脸或逼真图像等深度伪造功能，74.9% 的 NSFW 卖家于 2025 年入驻。这些卖家多使用 Stable Diffusion 和 ComfyUI 等开源工具，超半数服务提及 OnlyFans，美国与英国卖家占比超 25%。研究建议平台采用关键词检测、要求定制模型训练获得同意验证，并联合支付处理方施压合规。
正文
当前显示原文，中文全文译文尚未就绪。
← Back to blog
This is a blog post about our paper "From Underground to Mainstream Marketplaces: Measuring AI-Enabled NSFW Deepfakes on Fiverr" (USEC 2026, co-located with NDSS), by Mohamed Moustafa Dawoud, Alejandro Cuevas, and Ram Sundara Raman. It is a measurement study of AI-generated NSFW content on Fiverr, a mainstream freelance platform with 3.5 million active buyers.
What we found
We identified roughly 600 gigs advertising AI-generated NSFW content on Fiverr, despite the platform's explicit prohibition. The growth is fast: 74.9% of NSFW sellers joined Fiverr in 2025, compared to just 28% in mainstream AI categories.
Other findings:
82.8% of listings offer deepfake-enabling features such as face-swapping or photorealistic imagery.
Over half the gigs mention OnlyFans, and Instagram–OnlyFans combinations are the most common. This suggests Fiverr works as upstream infrastructure for content that violates other platforms' policies.
NSFW sellers mostly use open-source tools like Stable Diffusion and ComfyUI.
Western sellers (US and UK) make up over 25% of this segment, even though both countries recently enacted deepfake legislation.
What we recommend
We argue that platforms should use predictable keyword-based detection, require consent verification for custom model training requests, and engage payment processors to apply pressure for compliance.
Applying to PhD programs, or thinking about it? I mentor a cohort each cycle.