Overview
This series explains why perceptual hashes are similarity signals rather than identity fingerprints. It moves from the perceptual contract and matching thresholds to population-dependent NeuralHash and PhotoDNA errors, then separates natural failures from adversarial evasion and collision attacks. CertPHash shows how bounded guarantees can be stated and verified, while the final engineering test connects evidence, scale, workflow, cost, and deployment consequences without claiming universal safety.
Learning outcomes
- A perceptual hash encodes an application-defined similarity neighborhood, not byte identity or cryptographic proof.
- NeuralHash and PhotoDNA show that population, distance rules, thresholds, and comparison volume shape errors beyond nominal output size.
- Natural false positives and false negatives differ from deliberate evasion and targeted collision attacks, so each needs separate measurement.
- CertPHash can provide scoped bounded-robustness claims, but certificates do not replace population testing or establish universal safety.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why is a perceptual hash not a cryptographic fingerprint?
A perceptual hash is not a cryptographic fingerprint because it intentionally keeps selected visual variants close instead of making every changed input unrelated. A match means only that two representations fall within an application-defined similarity neighborhood.
How do thresholds and image populations change matching errors?
Thresholds and image populations change matching errors by determining which representation neighborhoods are accepted and how often real inputs occupy them. Wider windows can recover more similar variants while admitting more distinct images, and observed rates do not automatically transfer between populations.
What can certified perceptual hashing actually guarantee?
Certified perceptual hashing can guarantee only the scoped property checked by its model, verifier, norm, radius, threshold, and tested inputs. It can certify bounded no-evasion or no-collision results without proving universal resistance or eliminating natural population collisions.
What should be measured before deploying a perceptual hash?
Before deployment, measure intended invariances, both natural error directions, representation behavior, thresholds, comparison volume, adversarial evasion and collision, workflow safeguards, cost, and version-specific coverage. The claim should remain limited to the populations and conditions actually tested.