The perceptual fingerprint, explained simply

To recognise content regardless of its name or format, platforms, stock libraries and copyright-protection tools rely on a perceptual fingerprint. Understanding how it’s computed tells you what makes it change, and what leaves it untouched.

Cryptographic versus perceptual fingerprints

A cryptographic hash such as SHA-256 changes completely as soon as a single byte of the file changes. It’s used to check that a file is intact, not to recognise content. A perceptual fingerprint does the opposite: two images that look alike must produce close fingerprints, even if their files don’t share a single byte.

The four-step calculation

The average hash is the easiest to understand. It’s the one Klypsed uses to measure its variants.

1. Shrink the image to 8 × 8 pixels. All fine detail disappears; only the overall structure remains.
2. Convert to greyscale. Colour is dropped and only brightness counts.
3. Compute the average of the 64 brightness values.
4. Write one bit per pixel: 1 if the pixel is brighter than the average, 0 otherwise. You get 64 bits.

Other methods refine the idea. dHash compares each pixel with its neighbour. pHash goes through a discrete cosine transform to keep only the image’s low frequencies. Platforms use even more robust variants: in 2019 Meta published PDQ for photos and TMK+PDQF for videos.

Comparing two fingerprints

Two fingerprints are compared with the Hamming distance: the number of bits that differ. Zero means near-identical images. The higher the distance, the further apart the content. Each system sets a threshold below which two pieces of content are treated as the same.

For a video, the calculation is repeated on key frames spread over time. Sound has its own fingerprint, built from the energy peaks of its spectrum, on the same principle as music recognition apps.

What moves the fingerprint

• A horizontal mirror swaps the position of light and dark areas: many bits change at once.
• Cropping and zooming shift the image’s structure within the 8 × 8 grid.
• Changes in light and contrast push some pixels to the other side of the average.
• Speed shifts key frames in time, and pitch moves the peaks of the audio spectrum.

What leaves it untouched

• the file name, its date and its metadata;
• the container format, MP4 or MOV;
• a new encode at equivalent quality;
• a slight change in resolution, since the image is shrunk to 8 × 8 anyway.

Why measure rather than guess

A change can be visible and still barely move the fingerprint, or the other way round. The only way to know is to measure. Klypsed computes the fingerprint of the original and of every variant, then shows how many of the 64 bits changed. For a mirrored variant, the distance is measured against the mirrored original, so the mirror alone doesn’t count.

To put this into practice, read our method for posting the same video on several accounts.