Passport Verification Online: How Automatic Verification Works in KYC

Stages of Automatic Passport Verification in KYC

Automatic passport verification is a chain of sequential steps, each of which solves a separate task: accepting images, extracting data, assessing document authenticity, matching biometrics, and verifying live presence. The API returns a positive result to the customer’s backend only when all stages have been passed successfully.

The order and set of steps vary: depending on regulatory requirements, document type, and business scenario, the platform makes it possible to enable or disable individual checks, change sensitivity thresholds, and add additional data sources. Below is a typical pipeline for online passport verification in a KYC scenario.

Collecting the Passport Photo and Selfie

Passport verification online begins with image capture. The user photographs the passport spread and takes a selfie through an SDK or web widget embedded in the customer’s application or website. On the client side, a module works that helps aim the camera at the document: it prompts the tilt angle, controls focus and lighting, and automatically captures the frame when it meets the minimum quality requirements.

The quality of the input data directly affects the accuracy of all subsequent steps. A blurry image, glare from the laminate, or a cropped edge of the page can lead to a recognition error or a false trigger of the anti-fraud module. The client SDK contains built-in prompts: a frame for aligning the document, a lighting indicator, and automatic re-capture when the frame quality is insufficient. For the selfie, the module checks that the face fully fits in the frame, that there is no obstruction by hands or foreign objects, and that there is enough light for biometric analysis.

SDKs are available for iOS, Android, and web, and server-side libraries for Python, Java, and C#. This makes it possible to integrate image capture into virtually any channel: a mobile application, a web account, a chatbot, a CRM, or a self-service terminal.

Determining the Document Type and Image Quality Control

After the image arrives on the server, the system automatically determines the document type: a Russian citizen’s passport, an international passport, an ID card, a driver’s license, or another document from the base of supported templates. Classification is performed by neural-network models trained on tens of thousands of document types from more than 200 countries in more than 90 languages.

In parallel, image quality control is launched. The algorithm assesses resolution, contrast, and the presence of glare, shadows, and distortions. If the image does not pass the minimum thresholds — the text is unreadable or the document was photographed at a critical angle — the system immediately returns a request for a repeat image with an indication of the specific reason for the rejection. The user receives feedback within fractions of a second rather than after passing the entire chain of checks — this reduces the number of false rejections and saves time.

Automatic determination of the document type is especially important for international scenarios, where a single service accepts passports, ID cards, and residence permits from different jurisdictions. The system itself selects the required field-extraction template without requiring the user to manually specify the country of issue or the document type.

Recognition of Passport Fields and the Machine-Readable Zone

AI-OCR extracts textual and graphic data from the image: last name, first name, patronymic, date of birth, passport series and number, date of issue, department code, place of birth, and the owner’s photograph. For a Russian passport, the registration address is additionally recognized — a nontrivial task, since these data are often entered in handwriting or with a stamp of variable print quality.

A separate module processes the machine-readable zone (MRZ) — two lines of 44 characters each at the bottom of the data page, printed in OCR-B font per the ICAO Doc 9303 standard. The MRZ contains the owner’s encoded personal data and five check digits calculated by the «modulo 10» algorithm with a 731 weighting function. The system recognizes the MRZ, automatically recomputes the checksums, and compares the result with the recognized values. A discrepancy in any check digit is a signal of a possible recognition error or data modification.

OCR results from the visual zone (VIZ) and from the MRZ are reconciled with each other: the name, date of birth, and document number must match. Cross-validation of two independent sources within a single document increases the reliability of the extracted data and reduces the likelihood that a recognition error or the substitution of one of the fields will go unnoticed. The extracted data is normalized into structured JSON and passed further down the pipeline.

Verifying Document Integrity and Signs of Forgery

Automatic verification of passport authenticity is one of the most critical stages. The anti-fraud module analyzes the visual and structural features of the document to identify physical or digital forgery. The range of checks is broad: from basic ones (the correspondence of the layout to the expected template, the presence and placement of mandatory elements) to specialized ones (analysis of microtexts, background grid patterns, MRZ font characteristics).

The algorithms assess whether the image has been subjected to digital processing: pasting in a photograph, retouching data, replacing individual fields, or overlaying text in a graphics editor. Detectors identify anomalies at the pixel level — lighting inhomogeneity, JPEG compression artifacts in the area of a suspected edit, a mismatch in perspective between the document’s background and a pasted-in element. Separately, it is verified that an original physical document was presented rather than a photograph of a screen, a printout, or a scanned copy.

Based on the results, a set of flags and a document trust score are formed. If the risk level exceeds the set threshold, the case is routed to manual review by an operator or rejected automatically — depending on the scenario settings. The platform employs more than 40 anti-fraud algorithms working in combination.

Matching the Selfie Against the Photo in the Passport

After the photograph is extracted from the document, biometric verification is launched — comparing the face in the selfie with the photo in the passport in 1:1 mode. The neural-network model builds a biometric template for each image and calculates the degree of similarity.

The difficulty of the task is that the passport photo and the selfie were taken under different conditions: the lighting, angle, facial expression, and the person’s age differ — years may pass between the issuance of the passport and the moment of the check. Modern face verification models take these factors into account and are resilient to natural changes in appearance: glasses, a beard, age-related changes. The NeuroVision algorithm, which has undergone NIST FRVT benchmarking, provides a verification accuracy of 99.74% (a benchmark per the developer’s data) at a comparison speed of under 0.1 second.

The result is a numerical match score and a binary match/no match decision. The threshold value is configured for the specific scenario: for bank onboarding, the acceptable level of false matches is stricter than for registration on an informational portal.

Verifying Live Presence in Front of the Camera

The liveness check (PAD — Presentation Attack Detection) answers the question: is there a live person in front of the camera, or is a photograph, video recording, 3D mask, or deepfake being presented? Without this step, the entire preceding chain loses its meaning — an attacker could substitute a high-quality image of someone else’s passport and a matching photo of its owner’s face.

A common approach is a passive liveness check: the user simply looks into the camera while the system analyzes micro-signals unavailable in a static image — micro-movements of the eyes, blinking, skin texture, scene depth, the nature of reflections, and lighting transitions. A passive check minimizes friction for the user: they do not need to perform special actions (head turns, pronouncing digits), which has a positive effect on conversion.

The result is pass/fail and a numerical score. The check takes under a second. The stated accuracy of the liveness module is 99.9% (developer’s benchmark). In the on-premises variant, the quality of the algorithms is identical to the cloud one — the models are delivered in Docker/VM containers and run within the customer’s secure environment.

The liveness check is used as a step within the KYC pipeline or as a standalone module — for example, during a user’s repeat authentication to confirm an operation.

Additional Cross-Checking of Passport Data Against External Databases and Customer Data

The final stage of passport verification by photo is cross-checking the extracted data against external sources. The set of checks depends on the scenario, the connected databases, and regulatory requirements.

Typical checks for Russian passports: validation of the series and number against the registry of invalid passports, obtaining the INN by full name and passport data, checking the correspondence of the full name and SNILS, confirming the «full name — phone number» or «full name — email» link. To assess financial risks, checks are connected against the Federal Bailiff Service (debts and enforcement proceedings), the bankruptcy registry, the Ministry of Internal Affairs wanted database, and the list of terrorists. If the KYC scenario includes compliance requirements, AML checks are launched in parallel: screening against sanctions lists (OFAC, EU, UN, national lists), PEP registries, and adverse media databases.

For legal entities and their representatives, a KYB loop is available: verifying the company against the Unified State Register of Legal Entities (EGRUL) and foreign registries, identifying the ownership structure and ultimate beneficial owners (UBO), and screening directors and founders.

The results of all checks are aggregated into a single report indicating the status of each source, the reasons for mismatches, and the final trust score. Based on the report, the customer’s backend system makes a decision automatically or passes the case to an operator if the result is ambiguous.

How the Verification Result Is Formed

After the passport has been photographed, the data recognized, the document checked for authenticity, and the selfie matched against the owner’s photo, the system forms the final decision. The result is a multi-layered structure of scores, flags, and reasons, on the basis of which the customer’s backend makes the final decision on the customer.

The mechanism for forming the result determines how quickly a user passes onboarding and how many applications require operator involvement. This is where the acceptable-risk thresholds, routing rules, and escalation logic are configured — what turns recognition technology into a manageable business process.

When a Decision Is Made Automatically

An automatic decision is made when all the checks in the chain have concluded with an unambiguous result: the document is recognized correctly, the data matches the MRZ, no signs of forgery have been detected, the selfie matched the photo in the passport, and the liveness check has been passed. In such a situation, the system assigns the application the status «verified» without operator involvement.

Technically, behind this is aggregated scoring: each check — OCR, document integrity, biometric comparison, anti-fraud algorithms — returns its own reliability score. If all scores exceed the set thresholds and no anti-fraud algorithm has recorded suspicious signals, the application passes automatically.

The thresholds are configured for the specific business scenario. For a microloan, one level of strictness is acceptable; for opening a brokerage account, another. The platform makes it possible to set parameters in the KYC scenario: which checks are mandatory, what minimum score is acceptable at each step, under which combinations of factors the application is considered safe.

In practice, the share of fully automatic decisions in mature KYC systems reaches 85–95% of the total flow — provided the input data is of high quality and the thresholds are correctly configured. The remaining applications are routed to a repeat capture or manual review.

When a Repeat Image or Manual Review Is Needed

Not every application that fails automatic verification means a rejection. The system distinguishes two types of situations: fixable ones and those requiring expert assessment.

Fixable problems are cases where the quality of the input data did not allow the check to be performed: a blurry passport photo, glare on the lamination, cropped edges of the document, insufficient lighting when taking the selfie. The API returns an instruction to the customer to request a repeat image from the user with a description of the specific reason: «glare on the document photo», «face partially covered», «document cropped». The user receives a clear instruction, takes a new frame, and the check is launched again without operator involvement.

Manual review is engaged when the automation detects an ambiguity but cannot confidently classify it as fraud. Typical scenarios: a borderline value of the biometric comparison (the score is close to the threshold but does not cross it), individual anti-fraud signals in the absence of a complete picture, a non-standard document format, visual anomalies that the algorithm recorded but could not unambiguously attribute to forgery.

The application enters the operator interface (back office) with the full set of collected data: document images, the selfie, the extracted fields, the results of each check, the score, and the list of triggered flags. The operator sees the specific reasons for the escalation and can make an informed decision — confirm, reject, or request additional documents.

What Data and Statuses the API Returns to the Customer’s System

The KYC platform interacts with the customer’s backend via REST API and returns a structured response in JSON format. The composition of the data depends on the set of connected modules, but a typical response includes several blocks.

Extracted document data: full name, date of birth, passport series and number, department code, date of issue, MRZ data. Along with the fields, quality metadata is transmitted: a readability score, the presence of glare or cropping, the type of recognized document.

Authenticity check results: document integrity status, the correspondence of the MRZ to the visual fields, detected anomalies and their description.

Biometrics: the result of comparing the selfie with the photo in the document (match/no match) and a numerical match score, the result of the liveness check (pass/fail) with its own score.

Anti-fraud signals: the list of triggered rules, the aggregated risk score, flags of suspicious activity.

When additional modules are connected, the response is expanded. The AML check adds the results of screening against sanctions lists and PEP databases. Checks against external databases return confirmation of the «full name — phone number» link, the INN status, Federal Bailiff Service data, and other information — depending on the connected sources.

All data is available both through a synchronous API response (typical time — under one second) and through webhook notifications for asynchronous scenarios. The full history of checks is saved in a log and available through the operator interface — for audit, review of disputed situations, and the generation of reporting for the regulator.