How KYC Systems Detect Forged Documents

How a KYC System Checks a Document

Document verification in KYC is a chain of sequential stages, each of which solves its own task. The system accepts the image, determines the document type, extracts the data, assesses authenticity, and forms the final verdict. The entire process takes seconds and proceeds without human involvement — until the algorithm detects an anomaly.

Document Capture and Quality Control

Verification begins with obtaining the image. The user photographs the document through the smartphone camera or uploads a scan, and the client-side SDK — a mobile or web application — transmits the file to the server. The quality of the input image is critical: the subsequent algorithms work with pixels, and if the image is blurry, cropped, or shot at a sharp angle, recognition accuracy drops sharply.

Therefore, even before sending to the server, the SDK performs preliminary filtering: it checks the sufficiency of resolution, the absence of critical overexposure and shadows, the presence of all corners of the document in the frame, and an acceptable level of perspective distortion. If the quality does not pass the threshold, the user receives a prompt — to re-photograph, remove glare, position the document more evenly. Control at the input reduces the share of erroneous rejections and saves server resources: only usable images go for processing.

Document Classification and AI-OCR

Before extracting data, the system determines what is in front of it: a passport, driver’s license, ID card, residence permit, or another document type. A classification model based on convolutional neural networks determines the type and country of issue by visual features — the placement of fields, the color scheme, the size and structure. Mature platforms support tens of thousands of templates from hundreds of countries and process documents from virtually any jurisdiction.

After classification, AI-OCR comes into play — optical character recognition based on neural networks. The technology recognizes printed and handwritten text, extracts the values of specific fields — first name, last name, date of birth, document number, expiry date — and forms a structured data set in machine-readable format. The recognition accuracy for printed fields in mature solutions reaches 99.8–99.9%; for handwritten text it is lower — around 97–98%, which is due to the variability of handwriting.

Classification and OCR are the foundation for all subsequent checks: without correctly extracted data it is impossible to reconcile fields, assess the template, or form the final score.

Verifying Document Authenticity

The extracted data and the image itself undergo a multi-level authenticity check. Each level targets a specific class of falsifications — from primitive text replacement to professional digital substitution of fragments.

Field and MRZ Reconciliation

The first control is the internal consistency of the data. The system reconciles the values extracted from the document’s visual zone (VIZ) with the data of the machine-readable zone (MRZ). The MRZ is a standardized area at the bottom of a passport or ID card whose structure is defined by the ICAO 9303 standard. It contains the owner’s name, citizenship, document number, date of birth, sex, and expiry date, encoded in OCR-B font in a strictly fixed format.

The key protection mechanism of the MRZ is check digits. They are computed by the «modulo 10» algorithm with cyclically repeating weighting coefficients of 7–3-1 for each of the critical fields: the document number, date of birth, expiry date. If even one character in a field is changed, the check digit stops matching. The system recomputes each checksum and compares the result with the one specified in the document. A mismatch is a signal of a possible forgery or mechanical damage.

In addition to the MRZ, all duplicated data is cross-reconciled: the name in the visual zone and in the MRZ, the date of birth on different pages, the document number in different places on the form. Any discrepancy is recorded and passed to the scoring stage.

Checking the Template and Layout

Each document type has a reference template: fixed field positions, font sizes, a color palette, the placement of the photograph, the emblem, and security features. The system reconciles the received image against the reference and records deviations.

The geometric proportions of the fields, the correspondence of fonts (typeface, point size, letter spacing), the placement and dimensions of the photograph, the color scheme of the background and layout elements are analyzed. A forged document made from an inaccurate template gives itself away even at the level of an indentation of fractions of a millimeter — neural-network models record discrepancies imperceptible to the human eye.

Separately, the logic of the content is assessed: the series and number must correspond to the format of the issuing country, the date of issue must fall within an acceptable range, the age in the photo must be coherent with the date of birth. These checks weed out crude forgeries created without knowledge of the layout rules of the specific document.

Checking Integrity and Security Features

At the next level, the system analyzes the physical signs of authenticity available in the digital image. Modern passports and ID cards contain security features, some of which remain visible in the photograph: microtext, guilloche patterns, optically variable elements (OVI), holograms, and kinegrams.

Neural-network models are trained to recognize the presence and correct placement of these elements. The absence of a hologram, its displacement, or its replacement with a solid fill is recorded as an anomaly. The integrity of the background patterns is assessed similarly: an interruption of a guilloche line in the area of the photograph or text may indicate a montage — the paste-in of a fragment from another source.

A separate block is the analysis of lamination integrity. On forged documents, traces of delamination, creases, or non-uniformity of reflection are often noticeable in the zone of data substitution. The algorithms assess the surface texture and the uniformity of glare to detect such traces.

Detecting Digital Image Substitution

The most complex class of forgeries is digital modification: an attacker takes a scan of a genuine document and edits individual fields or the photograph in a graphics editor. To detect such manipulations, digital image forensics is applied.

One of the basic methods is Error Level Analysis (ELA). The image is re-saved with a known level of JPEG compression, after which it is compared with the original pixel by pixel. Areas that have undergone editing and re-saving exhibit anomalous compression artifacts differing from the rest of the document.

In addition to ELA, supplementary methods are used. Metadata analysis (EXIF) makes it possible to detect traces of processing: a mismatch between the declared camera and the actual resolution, the presence of graphics-editor markers. Models based on convolutional neural networks assess the uniformity of the image’s noise structure — for a fragment inserted from another file, the noise profile inevitably differs from the document’s background. Individual algorithms detect cloning — copying a section of an image within the same file to mask deleted data.

The combination of these methods forms a multi-layered barrier: even if a forgery passes one check, another will likely catch it.

Risk Scoring and Manual Review

The results of all checks are aggregated into a single risk score. Each signal — a check-digit mismatch, an anomalous font, a suspicious error level in the photo area — is assigned a weight, and based on the results a numerical trust score for the document is formed.

Based on the score, the system makes one of three decisions: the document is recognized as genuine and passes automatically; the document is rejected as a forgery; the document is routed to manual review by an operator. The thresholds for each decision are configured for the specific business process: a bank with strict regulatory requirements will set a lower automatic-approval threshold, while an online service with mass registration will set a higher one, so as not to overload operators.

Manual review is a deliberate element of the architecture, not a sign of the system’s weakness. It is needed for borderline cases: the document is damaged but genuine; the image quality is on the edge of acceptable; the document format is atypical and not fully covered by the template. The operator sees the document image, the results of each automatic check with explanations, the extracted data, and the final score — and makes a considered decision within minutes.

Mature KYC platforms automatically pass up to 90% of applications without human involvement. The remaining 10% are not necessarily forgeries but largely non-standard cases for which the algorithm lacks confidence. This approach balances the speed of passage for legitimate users against the reliability of detecting falsifications.

Which Signals Beyond the Document Help Detect a Forgery

Checking the document itself is only part of the equation. Even a technically flawless forgery fails if the system analyzes the context: whether the face in the selfie matches the photo in the passport, whether there is a live person in front of the camera, whether the device parameters and the user’s behavior point to a fraudulent scheme. This external verification loop turns KYC from a document scanner into a full-fledged anti-fraud barrier.

Matching the Selfie Against the Photo in the Document

After extracting the photograph from the document, the system compares it with a selfie the user takes in real time. Technically this is a 1:1 biometric verification task — the algorithm builds a mathematical representation of the face (a biometric template) for each image and computes the degree of similarity between them. If the score is below the threshold value, verification is not passed.

The main challenge is the difference in shooting conditions. The passport photo was taken in a studio under controlled lighting, possibly several years ago. The selfie is shot on a smartphone camera, often under uneven light, at an arbitrary angle, and the person’s appearance may have changed: a beard, glasses, a different hairstyle appeared, weight changed. Modern neural-network models take these variations into account and compare not pixels but the geometry and unique features of the face — the distances between key points, the shape of the cheekbones, the contour of the brow ridges — which makes it possible to work correctly even with substantial age-related and visual differences.

The accuracy of comparison determines two critical indicators: the false rejection rate (when the real owner of the document is denied verification) and the false match rate (when a fraudster passes on someone else’s document). Algorithms tested by the NIST FRVT methodology demonstrate verification at the level of 99.7–99.9% with a false match rate on the order of one case per million comparisons. For this task, the NeuroVision platform uses the NeuroVision module, which in NIST FRVT testing is among the best solutions in the world (within the TOP-30) and holds the first position among Russian participants.

A nuance that is often underestimated is the quality of the input image. If the client-side SDK does not control sharpness, lighting, and face position before the selfie is sent to the server, the algorithm receives a noisy input, and accuracy drops regardless of the power of the model. That is why mature KYC solutions include a quality-control module directly on the user’s device — this reduces the number of repeat attempts and speeds up conversion.

The Liveness Check Against Photos, Video, and Deepfakes

A face match by itself does not guarantee that a live person is in front of the camera. A fraudster can hold a printed photo of the document owner up to the camera, play back a video recording on a tablet screen, or — in more advanced cases — apply a real-time deepfake, substituting their own face with a synthetic image of the victim.

To counter such attacks, the KYC pipeline uses a liveness module, or, in the terminology of the ISO/IEC 30107 standard, Presentation Attack Detection (PAD). Its task is to confirm that the biometric sample comes from a live, physically present person rather than from an artifact.

There are two approaches. Active liveness requires the user to perform an action — blink, turn the head, pronounce a phrase. Passive liveness analyzes a single frame or a short fragment of the video stream without additional actions on the user’s part: the system assesses micro-movements of the eyes, skin texture, the nature of reflections, scene depth, and dozens of other features that distinguish a live face from a flat image or digital generation.

The passive approach is more convenient and does not increase the verification time — the result is returned in under a second.

The threat of deepfakes has qualitatively changed the attack landscape. According to industry research, the number of attempts to bypass biometric checks using deepfakes and real-time face swap has grown several times over the past two years. A separate vector has become injection attacks: the attacker does not show the forgery to the camera but injects the synthetic image directly into the software data-transmission channel, bypassing the physical sensor. The European specification CEN/TS 18099 already defines requirements for detecting such attacks (Injection Attack Detection, IAD), and at the ISO level the global standard ISO/IEC 25456 is under development, which will formalize procedures for testing resilience to injections.

An industry benchmark for the reliability of a liveness module is a conformity assessment against ISO/IEC 30107-3 through accredited laboratories. Testing is conducted at several levels of attack complexity: at the first level, protection against simple attacks (printed photos, video on a screen) is checked; at the second, against more complex ones (high-quality 3D masks, composite dummies); at the third, against expert attacks using specially made artifacts. When choosing a solution, it is worth clarifying which level was passed and in which laboratory, as well as requesting the APCER (the share of successful attacks) and BPCER (the share of false rejections of genuine users) metrics.

A reliable liveness module is a critical element of a KYC system. Without it, biometric verification turns into a formality: a fraudster who has obtained a photo of the victim from open sources overcomes the check in seconds.