The Complete Guide to False Positives in Image Forensics

The most dangerous output of any forensic tool is a confident false positive, because it produces a specific, checkable-sounding accusation that happens to be wrong.

Here is a catalogue of the ways genuine photographs fail our tests.

Modern smartphone processing

This is the largest category. A current flagship phone photo has been through multi-frame stacking, machine-learned denoising, local tone mapping, sharpening and semantic segmentation before you ever see it. The result is a genuine photograph of a real scene whose noise statistics look nothing like a raw sensor capture.

Such images routinely score low on the noise layer. They are not fake. They are processed.

Platform re-compression

Every upload to a social platform re-encodes the image, often more than once. By the time you download it, you may be analysing the fourth or fifth compression generation. Compression and ELA signals are substantially degraded, and the composite score drifts down.

Legitimate editorial work

Cropping, straightening, exposure adjustment, colour grading, dust spotting and noise reduction are normal practice in every publication on earth. They alter pixels. Forensics detects that alteration and cannot judge intent.

Screenshots

A screenshot flattens all prior history into one new capture, at the screen’s colour profile and resolution. It also frequently introduces scaling. Almost every forensic signal is destroyed or replaced.

Format conversion

PNG to JPG, HEIC to JPG, WebP to PNG — each conversion rewrites the compression history. A PNG has no JPEG artefacts at all, which can read as anomalous.

Extreme aspect ratios and crops

Our metadata layer rewards standard aspect ratios because generators frequently emit at native model dimensions. A legitimately cropped panorama or a deliberately square editorial crop scores lower for entirely innocent reasons.

Low-light and high-ISO originals

Heavy sensor noise pushes the noise layer in the opposite direction — a genuinely noisy image can score unusually high, masking manipulation. False negatives have causes too.

Synthetic-but-honest imagery

An illustration, a 3D render, a diagram or an intentionally AI-generated image used openly as an illustration will all read as “manipulated,” because they are. The tool measures artefacts, not honesty.

The discipline this implies

Before treating a low score as meaningful, rule out every item above:

1. Is this the original file, or a platform copy?
2. Was it shot on a modern smartphone?
3. Has it been converted between formats?
4. Is it a screenshot?
5. Was it cropped or routinely edited?

If you cannot rule these out, your finding is “inconclusive,” not “manipulated.” Publishing the stronger claim on the weaker evidence is how verification work damages its own credibility.