Oz Forensics

There is no product here other than the models. Liveness detection, face matching, deepfake resistance, injection attack detection and document character recognition are the entire offering, sold as software development kits and licences to organisations that embed them. The company does not sell a workflow, a case manager, a screening database or a compliance platform, so nothing survives the removal test. This is the purest form of the pattern in this lane, and it is why other identity verification vendors appear on the customer list rather than only banks.

Liveness detection produces a binary determination and nothing published describes what surrounds it. There is no documented confidence threshold an institution can set, no described escalation route when a check fails, no manual review or adjudication surface, and no guidance on how many attempts a person is permitted before a permanent decision attaches.

Configurability exists in the form of active or passive modes and single or multiple frame capture, but those are capture settings rather than oversight controls. The gap matters more for a component vendor than it might appear, because the institution consuming the signal inherits an outcome it did not produce and has no published basis for deciding when to override it.

This is the strongest showing on the axis in the index and it is held at the middle grade deliberately, because the evidence for the deciding property is ambiguous rather than present. External measurement is emphatic: two accredited laboratories, presentation attack detection at levels one, two and three, separate injection attack assessment, and submission of the face matching model to a national evaluation programme, with figures published for each.

What would lift this to the top grade is a mechanism by which the institution conducts its own validation. The trust centre does offer controlled access to documentation on request, which is closer than any peer has come, but the material located describes that documentation as security, privacy and policy material rather than model documentation, and an ambiguous claim is not credited.

Nothing describes drift monitoring, a validation summary, or a stated position on supporting a customer's model risk review under supervisory guidance. The published accuracy figures are also headline single numbers rather than error rates presented in both directions across conditions.

Both routes are satisfied. A named institution reports a quantified outcome at scale: Eurasian Bank onboarded one million clients using the biometrics. A second named bank deployment carries a quoted managing director for retail business development describing a procedure completed in under five seconds without requiring a mobile application install.

Independent parties with something at stake have assessed the technology repeatedly rather than once, across two accredited biometric testing laboratories and a national face recognition evaluation programme. The company also won a major interbank messaging network's innovation challenge in 2017. Deployment breadth is stated across more than twenty countries covering banks, fintechs, identity verification vendors, telecommunications and government bodies.

Both halves are answered, which is rare on this axis. Safety validation is the deepest located anywhere in the index: presentation attack detection testing at levels one and two by one accredited laboratory, separate presentation and injection attack detection assessments by a second accredited laboratory, and a level three presentation attack evaluation, which is a tier beyond what any peer in this lane holds.

Injection attack detection matters specifically because it addresses the newer threat where a manipulated stream is fed directly into the pipeline rather than shown to a camera. On stewardship, the on premises licensing option is a structural answer rather than an assurance, because an institution can run liveness and matching entirely inside its own environment so that biometric data never reaches the vendor at all, and on device processing narrows the exposure further even in the hosted model. The subprocessor list covers the remainder. No pooled dataset, shared negative list or cross customer fraud consortium was located, which distinguishes this vendor from two others in the same tier.

Better evidenced than most of the lane and grounded in architecture rather than assertion. The trust centre sets out privacy and data protection practices and names subprocessors, data protection regulation compliance is stated, and the on premises option means an institution can process facial biometrics without transferring them to the vendor at all, which answers the central privacy question for this product category structurally.

Short of the top grade because no privacy management system certification is held, no retention or deletion terms are published for facial images, liveness video or biometric templates in the hosted path, nothing addresses United States state biometric privacy law where several states impose separate consent and destruction duties with a private right of action, and no service provider position under United States financial privacy law is stated.

The first genuine trust centre encountered in this lane, launched in March 2026, and it is what the axis was written to reward. It carries the certifications and standards evaluations, the security measures and policies, the privacy and data protection practices, client information, staff training detail, controlled access to further documentation on request, and, most notably, a list of the subprocessors the company works with.

A published subprocessor list is the single item recorded as missing against every other vendor assessed in this sweep, including much larger ones. Information security management certification is held to the current 2022 edition and was achieved in December, not to a superseded edition, and a service organisation control report and data protection compliance are also stated. The remaining gap is that certificate numbers and audit periods are not published in the material located, so verification still depends on requesting documents rather than reading them.

Oz holds no financial licence and no financial regulator has assessed the product, which is unsurprising for a biometric component supplier. The credentials it does hold are technical conformity assessments against biometric testing standards rather than regulatory authorisations, and they should not be read across as regulatory standing.

Government bodies are claimed among the customer base across more than twenty countries, which would ordinarily imply passage through public procurement vetting, but no specific agency, authorisation or programme is named, so this rests on an unattributed claim rather than a citable admission. Product material is written against fraud and biometric attack rather than directly against customer identification or anti money laundering obligations, which sit with the platforms that embed it.

Submission of the face matching model to the national face recognition evaluation programme is the most meaningful fairness related step available in this field, because that programme measures demographic differentials by design and publishes results the vendor does not control.

Unlike a competitor in this same tier whose claim about that programme is ambiguous as to whose algorithms were tested, this one is unambiguous and independently corroborated by trade press reporting that the company's own face biometrics were benchmarked there. That places Oz ahead of most of the index.

It stops short of the top grade for the same reason as Incode: the company surfaces a single headline accuracy figure rather than the demographic breakdown its own submission generates, publishes no bias testing methodology of its own, and offers no analysis of how performance varies across the emerging markets that form the core of its footprint, where capture devices and lighting conditions differ sharply from evaluation conditions.

Nothing published reaches the person being checked, and the failure mode here is unusually invisible. A liveness rejection typically presents to the individual as a failed attempt with a prompt to try again rather than as a decision, so a person repeatedly refused by a model may never learn that a model was involved or that a determination was made about them.

The component position lengthens the chain further, because the signal passes from this vendor to an identity verification platform to the institution before it reaches the individual, and no allocation of responsibility across those parties is published. There is no accuracy guarantee, no remediation commitment, no appeal route and no stated right to have a failed capture reviewed by a person.

The strongest position on this axis found anywhere in the index, and it closes the exact gap recorded against the closest peer. Oz builds the models it sells rather than assembling components, which makes it the supplier rather than an intermediary and gives an institution a single accountable party.

That claim is independently corroborated because the models themselves have been submitted to three separate external evaluators, so the technology being tested is demonstrably the company's own. Critically, and uniquely in this sweep, the trust centre publishes the subprocessors the company is partnered with, which means a buyer can actually trace the chain beneath the vendor instead of being told it is short. What is still absent is a model or version identifier that an institution could record against an individual decision in order to reconstruct later which build produced it.

Integration is the product's whole delivery mechanism and the options are broad: web and mobile software development kits with front end customisation, on device and server side execution modes, hosted service and on premises licence, and configurable liveness modes covering active or passive checks and single or multiple frame capture.

The clearest evidence of integration depth is commercial rather than documentary, in that other identity verification vendors embed this technology inside their own platforms, which is a harder integration test than a direct customer deployment. A European distribution partnership extends reach.

It falls short of the top grade because nothing published describes named integration into core banking, account opening or origination platforms, and no partner directory, public status page or changelog was located.

The delivery choice is the widest in this lane and it is genuinely residency relevant rather than an assurance. The software is available as a hosted service or as an on premises licence, liveness can execute on the end user device or on a server, and the company states this flexibility exists specifically to meet differing regulatory requirements for personal data processing.

An on premises licence is the strongest possible residency answer because the biometric data need never leave the institution's own infrastructure. The trust centre's subprocessor list further tells a buyer who is involved in the hosted path.

It is held below the top grade because, for the hosted service, specific hosting regions and in country residency options are not enumerated in the material located, and no data transfer mechanism is described for customers outside the vendor's home jurisdiction.

No rates, tiers, volume bands, minimum commitments or trial terms were located. The company operates both a subscription model for its hosted service and a licensing model for on premises deployment, and while naming those two structures tells a buyer something about how a deal would be shaped, neither carries a published figure or unit. For a product normally priced per check or per licensed instance, neither the unit nor the band is disclosed, so a buyer cannot size a deployment without entering a sales process.

Reach is wide and deliberately weighted toward emerging markets, spanning more than twenty countries with deployments described across banks, fintechs, identity verification and know your customer vendors, telecommunications operators, gaming and government bodies, with stated depth across the Commonwealth of Independent States, Africa, Asia, Latin America and a launch in Indonesia. Named financial institutions are regional banks rather than global groups.

It sits below the top grade for a structural reason rather than a gap in effort: this is a component supplier whose route to many institutions runs through other identity verification vendors, so segment coverage is partly inherited from customers rather than held directly, and no deployment at a large multinational bank was located.