I’m particularly interested in specializing in the detection of AI-generated videos to distinguish them from authentic footage. With AI advancing toward near-perfect video replication, I see a significant risk in its potential for falsifying original content. I believe watermarking alone is insufficient, as it can be removed. Instead, I envision a detection method based on analyzing structural metadata: AI-generated videos typically start as text or image inputs, whereas authentic footage is recorded directly through a camera until any edits occur.
My idea involves embedding red-flag metadata tags, like “Edited with AI software (e.g., Sora)” when AI tools are used. Additionally, I’d propose pixel-level watermarks, hidden throughout the video at microscopic scales, undetectable by the human eye but offering traceability. This might involve invisible nanoscale watermarks across billions of microscopic details, coded across multiple axes (x, y, and z) to ensure even powerful systems couldn’t fully erase these embedded markers.