Current forensic tools only label images as real or fake, but this ignores important differences. For example, an AI-enhanced portrait that keeps the same people and objects is labeled fake, just like a photo with a completely fabricated scene. This can mislead journalists and moderators who need to know what actually changed and whether it matters.
The researchers propose 2CAP, a system that uses two parallel detectors. One detector checks if the image came from a camera (looking at sensor fingerprints), while the other checks if the scene content makes sense. By combining these two streams, the system can sort images into camera-captured, AI-retouched, locally tampered, or fully synthetic.
On a multi-source benchmark, 2CAP achieved the best classification performance, especially for retouched images where current methods fail. It also generates explanations that indicate what was altered. This allows content moderators and forensic analysts to base decisions on the type of modification rather than a single fake label.