Sensor Noise Is a Fingerprint. Here Is How to Read It

Every digital photograph is noisy. That noise is not a defect to be removed — it is evidence.

Where noise comes from

Photons arrive at a sensor at random intervals. That randomness alone produces shot noise, which follows a Poisson distribution and scales with the square root of brightness. On top of that sit read noise from the amplifier, thermal noise that increases with sensor temperature, and fixed-pattern noise unique to the individual sensor’s manufacturing variation.

The consequences matter for forensics:

• Noise is independent across colour channels, because each photosite measures a different wavelength band through its own filter.
• Noise is brightness-dependent — proportionally stronger in shadows.
• Noise is spatially uncorrelated at the pixel level, before demosaicing.

What generated images do instead

Diffusion models produce images by iteratively denoising. The final output has, by construction, had noise removed rather than added. What remains is residual texture that resembles noise superficially but has different statistics:

• It is often correlated across channels, because it emerges from a shared latent representation rather than three independent sensors.
• It frequently lacks the brightness dependence of shot noise.
• It can show spatial correlation introduced by upsampling layers.

Our noise layer measures the first of these directly: the mean absolute deviation between colour channels across the frame. Real photographs show substantial channel divergence. Generated and heavily denoised images show markedly less.

Why heavy editing looks similar

This is the honest caveat. Modern smartphone photography applies aggressive computational denoising — multi-frame stacking, machine-learned noise reduction, night modes that composite many exposures. The output is a genuine photograph whose noise has been substantially engineered away.

Such an image can score poorly on the noise layer despite being completely authentic. This is one of our most common false positives, and it is why the noise reading is never decisive on its own.

Practical guidance

• Always analyse the original file, not a messaging-app copy. Compression and re-encoding both attack noise statistics.
• Expect modern flagship phone photos to score lower on noise than older DSLR files. That is processing, not fakery.
• Treat a low noise score as one signal among four. Cross-reference with the ELA map before drawing conclusions.

The research frontier

The most powerful version of this technique is PRNU — photo-response non-uniformity — which extracts a sensor’s unique fixed-pattern fingerprint and can match a photo to the specific camera that took it. PRNU requires multiple reference images from the candidate camera, which puts it outside the scope of a single-image browser tool, but it is where serious device-attribution work happens.