False Claims, Credible Images: EpiReal-Bench and EpiReal-Attack

Evidence record

Author-reported 10,000-prompt benchmark and adversarial black-box image-generation evaluation on four commercial models.

Read the original source ↗

Claim-level citations (5)

• supportsCommercial image generators rendered false claims as plausible evidence in a benchmark: The paper presents constructed false claims and controlled model evaluations rather than verified deployment of these particular generated images in a real-world misinformation campaign.

  Abstract, benchmark and attack methodology

• supportsCommercial image generators rendered false claims as plausible evidence in a benchmark: The EpiReal authors describe a benchmark with 10,000 false-claim prompts and 10,000 generated images spanning ten claim categories and ten credible-looking visual formats.

  Abstract, EpiReal-Bench dataset composition

• supportsCommercial image generators rendered false claims as plausible evidence in a benchmark: The authors describe a tested mismatch in which a model may identify a claim as false when asked in text yet still generate an image that presents the claim as credible visual evidence.

  Abstract, first paragraph on text recognition versus evidence-like image generation

• supportsCommercial image generators rendered false claims as plausible evidence in a benchmark: The EpiReal authors evaluate verification-guided prompting, which asks an image generator to check a claim before drawing it. They report improved interception for most of the tested models without external retrieval, retraining or changes to model parameters, but do not present it as a guarantee of factual correctness.

  Full paper, Section 4.4 RQ3 comparison of defenses and Appendix C.4 Details for VGP

• supportsCommercial image generators rendered false claims as plausible evidence in a benchmark: In the author-reported experiments on four commercial image generators, over 70% of the false-claim prompts produced images faithfully depicting the misinformation; the authors report up to 95% with their EpiReal-Attack optimization.

  Abstract, Experiments on four commercial models sentence