Perceptual hash

How does it work?

The image is shrunk to a small grid, converted to greyscale, transformed into a few dozen numbers that describe its coarse structure (block averages, or the low frequencies of a discrete cosine transform), and those numbers are reduced to a string of bits. Two images are compared by counting the bits that differ (the Hamming distance). A re-saved JPEG or a resized copy differs by a few bits; an unrelated image differs by about half. Microsoft's PhotoDNA, Facebook's PDQ and the open pHash family are the common implementations.

How is it different from other hashes and from embeddings?

Where is it used?

• StopNCII.org and NCMEC's Take It Down compute the hash on the victim's device and share only the hash; participating platforms block uploads that match. The image never leaves the phone.
• Child sexual abuse material detection: PhotoDNA hashes of known material are matched at upload by most large platforms.
• Copyright and brand protection: finding reposts of a known image at scale.
• Reverse image engines such as TinEye combine hashing with other features to find copies and edits.

What are the limits?

Heavy crops, mirroring, rotation, overlays and screenshots of screenshots can push a copy past the matching threshold, and some systems add geometric variants to compensate. Hashes recognise only images already submitted; they cannot find new images of a person, which is the NCII gap that a face search partly covers and hashing does not. A hash cannot be reversed into the image, which is why sharing it is considered safe, though research has shown that collisions can be engineered for some algorithms.