Report Reveals Deepfake Vulnerabilities in Top Hugging Face Image Models

According to an investigation by the European non-profit organization AI Forensics, 7 out of the top 9 image editing models hosted on Hugging Face were found to carry out explicit deepfake requests, such as undressing photos of women.

Key characteristics and implications are as follows:

• Shift in Model Governance: While commercial API-based models generally block such requests by default, models that distribute open-source weights directly often lack adequate safeguards.
• Platform Accountability: Model hosting platforms were previously regarded as neutral distributors, but these findings increase the likelihood that platforms will become subject to regulation as new content moderation entities.
• Regulatory Environment: With deepfake prevention laws such as the Take It Down Act already in place in the US, the safety issues surrounding open-source models are expected to become an important issue for policymakers.