Forensic Knowledge Graphs authenticate images, beat VLMs and detectors

As generative AI produces increasingly realistic image forgeries, trustworthy authentication systems are critical. Existing forensic detectors target specific forgery types but lack interpretability, while vision-language models (VLMs) offer explanations but fail to reliably exploit forensic traces. To bridge this gap, Nguyen and Stamm propose Forensic Knowledge Graphs (FKGs), a unified framework that integrates forensic evidence extraction, structured reasoning, and human-interpretable explanation. FKGs encode traces, causal dependencies, and links to scene content, enabling both detection and justification of tampering.

The authors introduce a novel forensic authentication network to generate accurate FKGs, paired with an Iterative Context Refinement strategy that steers VLMs toward faithful, grounded explanations. They also release FKG-50K, a dataset of 50,000 realistic forgeries with ground-truth FKGs. In experiments, FKG outperforms both forensic detectors and VLMs across multiple metrics: detection accuracy, forgery identification, localization precision, and forensic justification quality. The work, accepted at ECCV 2026, marks a step toward explainable and reliable image forensics.