LightShed Presented at USENIX Security ’25

2025/09/22

LightShed addresses a growing concern in the age of generative AI: protecting digital artwork from being used without consent. As generative models are often trained on large datasets scraped from the web, artists worry that their work can be copied or their distinctive styles imitated. Tools like Glaze and NightShade attempt to counter this by introducing imperceptible perturbations into images, making them unusable for model training. In our work, we show that these tools offer only a false sense of protection. LightShed can automatically detect and remove such perturbations, exposing fundamental weaknesses in current image-based copyright defenses. We presented these findings at USENIX Security ’25 and were grateful for the insightful discussions, valuable feedback, and strong interest from the broader security and machine learning communities.