Reproduce “Toward Accountable AI-Generated Content on Social Platforms: Steganographic Attribution and Multimodal Harm Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinlei Guan, David Arosemena, Tejaswi Dhandu, Kuan Huang, Meng Xu, Miles Q. Li, Bingyu Shen, Ruiyang Qin and 2 more
A framework combines steganographic attribution and multimodal harm detection to address accountability for AI-generated content on social platforms. The system embeds cryptographically signed identifiers into images and utilizes a vision-language model to detect harm arising from image-text context, achieving robust detection with an AUC-ROC of 0.99 and enabling forensic tracing even after common image transformations.
Key insights
- Combining distinct multimodal semantic features (raw embeddings, difference, product, cosine similarity) into a fused vector significantly enhances the detection of contextual harm in image-text pairs.
- Spread-spectrum watermarking methods (Spatial SS, DWT-SS) offer superior robustness for embedding attribution information compared to bit-level embedding under common social media transformations.
- Asymmetric encryption provides a verifiable cryptographic identity layer, allowing for authentication of the embedded user identifier and protection against tampering for accountability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.