LETHE: Defending against Unauthorized Image-to-Video Generation via Value Diversity Collapse

Key points

• A promising solution is to proactively add an adversarial perturbation to the image before release, such that videos generated from it would be unfaithful or visually corrupted.1
• Through a convex geometric analysis, we theoretically identify the spread of the reference value vectors as a key factor governing this reading process, and empirically find that even a mild contraction of this spread can severely degrade the generated video, making it an efficient lever for protection.1
• Guided by this insight, we propose LETHE, a novel protection method that directly disrupts this reading process.1
• We further construct two human-centric I2V protection benchmarks and conduct extensive experiments with multiple modern I2V models, which demonstrate the state-of-the-art protection performance and versatility of LETHE.1

Sources (1)

• [1]LETHE: Defending against Unauthorized Image-to-Video Generation via Value Diversity Collapse
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 10:08 AM

  Image-to-video (I2V) diffusion models can animate a single reference image into high-quality, subject-consistent videos, posing threats of misuse as anyone can turn personal photos into fabricated videos without consent.

  A promising solution is to proactively add an adversarial perturbation to the image before release, such that videos generated from it would be unfaithful or visually corrupted.

Extractive summary: sentences quoted from the sources.