Challenges in authenticating synthetic media: forensic insights from 6,000 DALL·E-generated images

Research context

This paper examines noise-residue characteristics in camera and synthetic imagery, comparing balanced samples from the Dresden Image Database and DALL·E. The published abstract describes statistical modelling and likelihood ratios for evaluating competing origin hypotheses.

The study is part of the research foundation for SPOT. Results from this experiment should not be treated as a general performance guarantee across all cameras, generators, editing processes or operational conditions.

The publisher provides the abstract and access options for the full paper.

This RHEM Labs paper forms part of the ANZFSS Melbourne 2025 journal supplement, published in 2026.

Reported results and limitations

The Results section reports mean ROC AUC 0.998 and mean accuracy 0.985 over 100 randomly sampled runs. The abstract gives a different accuracy figure (0.998). The SPOT evidence section separates these figures and reproduces the counts from one reported test example. Results are specific to the described camera and DALL·E populations; they are not a general claim about current SPOT operational performance.