Face recognition accuracy in KYC: comparing biometric algorithms and face recognition errors

How face recognition accuracy is determined in KYC

Face recognition accuracy in KYC processes is determined by the algorithm’s ability to correctly match a person’s biometric data with their reference image or to correctly distinguish them from other people in the database. Unlike the general application of recognition technologies, where a certain margin of error is acceptable, KYC requires exceptional accuracy — an error can lead to financial losses, a violation of regulatory requirements, or fraud.

Accuracy assessment is built on the mathematical analysis of the probability of the algorithm making the correct decision when comparing biometric vectors. Each face is converted into a unique set of numeric characteristics — a biometric template of between 128 and 2048 dimensions. Comparing two templates gives a numeric similarity value, which is matched against the established decision threshold. It is the choice of this threshold that determines the balance between the strictness of the check and convenience for legitimate users.

In real KYC systems, accuracy is measured not by abstract percentages but by concrete operational indicators: how many fraudsters the system will let through per million checks, what percentage of genuine customers will encounter a false rejection, and how fast verification happens. Modern algorithms achieve an accuracy of 99.74% at a threshold ensuring just one error per million operations — a critical level for financial organizations.

1:1 verification and 1:N identification in KYC scenarios

1:1 verification checks whether a person really is who they claim to be. The algorithm compares the customer’s selfie with the photograph from the identity document. This is the basic scenario when onboarding new users: the system receives two images and must determine whether they belong to the same person. The similarity threshold is configured to minimize the probability of identity substitution while maintaining a comfortable process for real customers.

1:N identification solves a fundamentally different task — finding a specific person among many records in a database. The system compares the new image with all the stored templates and determines the most similar ones. In KYC, this mode is used to check for duplicate accounts, search sanctions lists with photographs, and detect attempts at repeated registration after a block. The complexity grows in proportion to the size of the database: with a million records, the probability of a random match increases a millionfold compared with a single verification.

For KYC, it is critical to understand the difference in the accuracy requirements of these modes. With 1:1 verification, a stricter threshold can be allowed — the customer is interested in passing the check and is ready to take a repeat selfie if necessary. With 1:N identification, an overly strict threshold will lead to an avalanche of false positives, especially in large databases. Therefore, systems use adaptive thresholds: softer ones for the initial search for similar faces and stricter ones for the final confirmation of a match.

Hybrid scenarios combine both approaches. When logging into a personal account, 1:1 verification against the user’s saved template is performed, but in parallel a 1:N check against a database of blocked or suspicious accounts is launched. This approach provides multi-level protection: it blocks both direct attempts at unauthorized access and the circumvention of restrictions through the creation of new accounts.

Which biometric accuracy indicators are critical specifically for KYC

For KYC processes, the general “recognition accuracy” metric is not enough. Financial organizations and regulators assess specific operational risks, each of which has direct financial consequences. The key indicator is the probability of letting a fraudster through at a given level of rejection of legitimate customers.

Processing speed directly affects onboarding conversion. If verification takes more than 3 seconds, up to 15% of users abandon registration. Modern systems process a biometric comparison in 100-200 milliseconds, but the total check time includes uploading the images, their preprocessing, and additional authenticity checks. Each additional second of waiting reduces conversion by 2-4%.

Robustness to changes in appearance determines how long the system will recognize the customer without repeated verification. The algorithm must work correctly with changes in hairstyle, the growing or shaving of a beard, age-related changes, and different makeup. The critical period is 5 years, after which accuracy begins to noticeably decline. Financial organizations update biometric data every 2-3 years or after substantial transactions.

The ability to work with varying image quality is critical for mass application. Customers use devices with cameras from 0.3 to 108 megapixels, under various lighting, with different shooting angles. The system must maintain acceptable accuracy even with compressed low-resolution JPEG images, partially obscured faces, and glare from glasses. The drop in accuracy when moving from ideal studio photos to real selfies must not exceed 5-7%.

Protection against presentation attacks — the ability to distinguish a live person from a photograph, video or mask. For KYC this is a critically important indicator, measured separately from the basic recognition accuracy. Liveness check systems must block 99.9% of deception attempts while maintaining convenience for real users. The integration of anti-spoofing into the recognition process must not increase the check time by more than 1-2 seconds.

Metrics for assessing the accuracy and quality of face recognition algorithms

Assessing the quality of biometric algorithms requires an understanding of specific metrics that differ from the standard accuracy indicators in machine learning. In the context of KYC processes, the correct interpretation of these metrics determines the balance between the security of the system and convenience for legitimate users.

Basic face recognition errors: FAR, FRR, FMR, FNMR

Biometric systems operate with four fundamental types of errors, each of which is critically important for business processes.

Category	Description
FAR (False Acceptance Rate)	The share of cases where the system mistakenly accepts an outsider as a registered user. In the KYC context, this means letting a fraudster through verification. At FAR = 0.01%, the system will incorrectly confirm the identity of one attacker out of 10,000 unauthorized access attempts.
FRR (False Rejection Rate)	The percentage of situations where the system rejects a legitimate user. A customer with correct documents fails the check, which leads to lost conversion and a worse user experience. FRR = 1% means that every hundredth honest customer will encounter problems with identification.
FMR (False Match Rate) and FNMR (False Non-Match Rate)	Are algorithmic metrics at the level of comparing biometric templates, unlike FAR and FRR, which are measured at the transaction level. FMR shows the probability that two different biometric samples will be mistakenly recognized as matching. FNMR is the probability that the system will consider two samples of the same person to be different.

The difference between FAR/FRR and FMR/FNMR becomes significant with multiple authentication attempts. If a user can make three login attempts, then the system’s FAR will be higher than the algorithm’s base FMR, since the attacker gets several chances at a successful attack.

How FAR/FRR and “accuracy in percent” are related

The marketing phrase “recognition accuracy of 99.9%” is often misleading, since it does not reveal the real ratio of error types. This figure can mean different things depending on the context of measurement.

At a decision threshold ensuring FAR = 0.1% and FRR = 0.1%, the overall accuracy will indeed be 99.9%. However, the same algorithm at another threshold may show FAR = 0.001% with FRR = 5%, which gives an overall accuracy of 97.5% with substantially higher security.

A correct assessment requires specifying both metrics. A system with “99% accuracy” may have FAR = 1% (critical for security) or FRR = 1% (conversion problems). For KYC applications, a low FAR is usually more of a priority even at the cost of an increased FRR, since letting a fraudster through costs more than rejecting a legitimate customer, who can be offered an alternative verification method.

Summary quality metrics: EER, ROC/DET curves, AUC

Integral indicators make it possible to assess the quality of an algorithm independently of the chosen decision threshold.

EER (Equal Error Rate) is the point where FAR equals FRR. This indicator is convenient for the quick comparison of algorithms: a system with EER = 0.5% is objectively better than a system with EER = 2%. However, EER does not always reflect real performance in production, where thresholds are configured for specific business requirements rather than for the equality of errors.

The ROC curve (Receiver Operating Characteristic) visualizes the ratio between the true positive rate and the false positive rate at all possible thresholds. The closer the curve is to the upper-left corner of the graph, the better the algorithm. ROC analysis makes it possible to choose the optimal operating point for the specific requirements of the business.

The DET curve (Detection Error Tradeoff) represents the relationship between FRR and FAR on a logarithmic scale. Unlike ROC, the DET curve better shows the differences between algorithms in the region of low error values, which is critical for high-load KYC systems. On the DET graph, the ideal algorithm tends toward the origin of coordinates.

AUC (Area Under Curve) is the area under the ROC curve, an integral metric from 0 to 1. Modern biometric systems demonstrate AUC > 0.99, but even a difference of 0.001 can mean a substantial improvement in real operating conditions. AUC = 0.9995 versus 0.9985 is a tenfold difference in the number of errors at a fixed threshold.

Choosing the working threshold by metrics for the KYC task

Configuring the decision threshold determines the practical effectiveness of the face recognition system in a specific business context.

For financial organizations with strict regulatory requirements, the typical choice is a threshold ensuring FAR ≤ 0.01% (1:10,000). This guarantees a minimal risk of letting fraudsters through, although the FRR may reach 1-3%. Rejected users are offered additional verification via a video call or an office visit.

Online services with a focus on the user experience often choose thresholds with FRR < 0.5% at a FAR of around 0.1%. This ensures smooth onboarding for 99.5% of customers with an acceptable level of security.

A cascade strategy involves the use of several thresholds. An initial check with a liberal threshold (FRR = 0.1%, FAR = 1%) quickly lets obviously legitimate users through. Doubtful cases undergo an in-depth check with a conservative threshold or are sent for manual moderation.

Dynamic adaptation of thresholds takes the context of the transaction into account. For high-risk operations (large transfers, changes of details), a strict threshold with FAR < 0.001% is set. For logging into a personal account without financial operations, a softer threshold is acceptable.

Comparing biometric face recognition algorithms for KYC

The choice of a face recognition algorithm for KYC processes determines not only the security of customer identification but also conversion, operating costs, and compliance with regulatory requirements. At the same time, a correct comparison of biometric systems requires an understanding of many nuances that go far beyond vendors’ advertising promises.

Why a single “face recognition accuracy” figure is not enough

Claims of a recognition accuracy of 99.7% or 99.9% say practically nothing about the real applicability of an algorithm in KYC scenarios. This figure can mean the share of correct decisions on a specific test dataset, accuracy at a certain similarity threshold, the result on ideal studio photographs, or even the theoretical maximum of the algorithm in laboratory conditions.

The real performance of a biometric system is determined by a set of metrics under various conditions. An algorithm with a stated accuracy of 99.8% may show a FAR (False Accept Rate) of 1% — that is, letting one fraudster out of a hundred through, which is critical for financial services. Or it may demonstrate an FRR (False Reject Rate) of 5%, rejecting every twentieth legitimate customer, which destroys onboarding conversion.

A single accuracy figure hides the algorithm’s behavior across different demographic groups. A system may show excellent results on middle-aged Caucasian faces but produce an unacceptable number of errors on Asian faces or people over 60. For an international business or services with a diverse audience, such unevenness becomes a critical limitation.

The context of application radically changes the requirements for an algorithm. Verification at onboarding with a quality selfie and document requires one set of characteristics, re-authentication through a laptop webcam another, and recognition in a video stream from a mobile device a third. An algorithm optimized for one scenario may be unsuitable for another, despite an impressive “overall accuracy”.

Independent benchmarks (NIST FRVT and others) for comparing algorithms

The NIST Face Recognition Vendor Test (FRVT) remains the gold standard for evaluating biometric algorithms. The US National Institute of Standards and Technology conducts continuous testing on datasets that include millions of images of various quality: from ideal passport photos to surveillance camera footage. The results are published openly with detail by data type, demographic group, and shooting conditions.

The NIST methodology rules out the possibility of tuning an algorithm to the test. Vendors provide compiled libraries without access to the test data. Testing is conducted on the institute’s isolated servers, which guarantees identical conditions for all participants and the impossibility of manipulating the results.

FRVT includes several tracks relevant to KYC applications. The Visa Photos track assesses work with documents and quality portraits. Wild Photos tests algorithms on uncontrolled images close to real user selfies. The Mugshot track checks the matching of photographs with a difference of several years, which is critical for the long-term verification of customers.

In addition to NIST, there are regional and industry benchmarks. The European FRONTEX project evaluates algorithms for border control with an emphasis on working with biometric passports. The Chinese MegaFace Challenge focuses on scalability — searching in databases of millions of faces. The academic LFW (Labeled Faces in the Wild) remains the base test for scientific publications, although modern algorithms reach the accuracy ceiling on it.

When interpreting benchmark results, it is important to take their limitations into account. NIST tests algorithms, not comprehensive systems with a liveness check and document processing. The datasets may not reflect the specifics of a particular region or the demographics of the target audience. Results on static images do not always correlate with performance on a video stream.

How to correctly compare algorithms on your own KYC data

An objective comparison of biometric algorithms on your own data requires a methodical approach and an understanding of statistical principles. The results of internal testing often diverge from public benchmarks due to the specifics of business processes, the features of the audience, and the technical limitations of the infrastructure.

Preparing the dataset and labeling for face recognition tests

The quality of the test dataset determines the reliability of all subsequent conclusions. The minimum volume for statistically significant results is 10,000 unique identities with several images each. To assess demographic robustness, a balanced representation of age groups, sexes, and ethnic types is required, corresponding to the real structure of the customer base.

Data collection must reproduce production conditions as closely as possible. If customers pass verification through a mobile application, the test selfies are collected through a similar interface with the same requirements for lighting and positioning. Documents are scanned or photographed in conditions typical for users, including glare, shadows, and imperfect focus.

Labeling requires a strict protocol. Each image is linked to a unique identity identifier, and the source (document, selfie, video frame), timestamp, shooting conditions, and demographic attributes are recorded. Particular attention is paid to borderline cases: twins, people after plastic surgery, images with a substantial difference in time.

Checking for duplicates and labeling errors is critically important. Even a small percentage of incorrectly labeled data can distort the results of the algorithm comparison. Double-blind labeling with subsequent cross-validation of disputed cases is recommended.

Train/test splitting and excluding identity leaks

Correct data splitting prevents overfitting and guarantees a fair comparison. The classic proportion is 70% for training and threshold tuning, 30% for final testing. At the same time, the split is done by identity, not by image: all photographs of one person must fall either into the training or the test set.

A temporal split more realistically reflects production scenarios. Data from the first 9 months is used for tuning, the last 3 months for evaluation. This makes it possible to account for data drift, changes in camera quality, and the evolution of user behavior.

Identity leakage between the sets is a common error leading to inflated metrics. The problem is aggravated when using public datasets for pre-training algorithms. A check for the intersection of the test data with the training sets of all the systems being compared is necessary.

Stratification ensures the representativeness of the sets. The distribution by age, sex, ethnicity, and image quality must be the same in the training and test parts. For rare categories (for example, people over 80), separate validation with an enlarged sample is applied.

Test scenarios: onboarding, repeat login, fraud cases

Onboarding new customers involves matching a selfie with the photograph from the document. Testing includes genuine pairs (the selfie and document of one person) and impostor pairs (attempts to use someone else’s document). The metrics are assessed at various similarity thresholds, modeling the balance between security and conversion.

The repeat login scenario checks the matching of a new selfie with the saved reference. Here robustness to changes in appearance is critical: a different hairstyle, glasses, beard, makeup. Testing is conducted on pairs of images with various time intervals — from a few days to several years.

Fraud scenarios assess resistance to attacks. Presentation attacks include attempts at deception through a photograph, a video on a screen, a mask, or a deepfake. Morphing attacks check the ability to detect synthetic documents with the combined features of several people. Testing requires specialized datasets with real deception attempts.

Cross-channel verification tests the matching of images from different sources: a high-resolution document versus a blurred selfie, a professional photo versus a webcam, daylight versus artificial light. This is the scenario closest to reality, revealing the practical limitations of the algorithms.

Performance under load is often ignored in comparisons, although it can become a bottleneck. Throughput, latency under parallel requests, and memory and processor consumption are measured. For cloud solutions, stability under peak loads and degradation with network problems are critical.

Face recognition errors in KYC: types and impact on customer verification

Even the most advanced face recognition algorithms inevitably make errors. In the context of KYC processes, each such error directly affects business indicators: from the loss of customers to financial risks and reputational losses. Understanding the nature of these errors and their consequences makes it possible to build an optimal verification strategy that balances security and user convenience.

False matches (False Match): risks to security and fraud

A false match occurs when the system mistakenly accepts an attacker’s face as the face of a legitimate customer. In terms of metrics, this is called the False Acceptance Rate (FAR) or False Match Rate (FMR). For KYC processes, this is the most critical type of error from a security standpoint.

At a FAR of 0.01% (one case per 10,000 attempts), a fraudster has a chance to gain access to someone else’s account using their own face or a quality forgery. On the scale of a large fintech platform with millions of transactions, even such a minimal percentage of errors can turn into substantial losses. Each false match potentially means unauthorized access to bank accounts, the theft of personal data, or the conduct of fraudulent operations in the customer’s name.

The consequences go beyond direct financial losses. Regulatory fines for violating data protection requirements can reach millions of rubles. The reputational damage from public cases of account hacking through biometrics takes years to repair. Customers lose trust in the platform, move to competitors, and spread negative reviews.

Targeted attacks using deepfake technologies and 3D masks pose a particular danger. Modern methods of generating synthetic faces are able to deceive basic recognition algorithms. Therefore, the integration of a Liveness check and multi-factor authentication, which substantially complicate the task for fraudsters, is critically important.

False non-matches (False Non-Match): impact on conversion and UX

A false non-match is a situation where the system does not recognize a legitimate user and blocks their access. The False Rejection Rate (FRR) or False Non-Match Rate (FNMR) metric reflects the frequency of such cases. Although this error is less critical from a security standpoint, its impact on the business can be devastating.

At an FRR of 1%, every hundredth customer will not be able to pass verification on the first attempt. For onboarding new users, this means a direct loss of conversion. Studies show that after three unsuccessful biometric verification attempts, 40% of users abandon registration. In e-commerce, each additional second of delay when placing an order reduces the probability of purchase by 7%.

The psychological effect of being denied access reinforces the negative perception. A customer whom their own banking system “did not recognize” experiences frustration and anxiety. They are forced to contact support, go through alternative verification procedures, and waste time. The accumulated irritation leads to the churn of even loyal customers.

False non-matches especially often arise when a user’s appearance changes: a new hairstyle, glasses, makeup, a beard, the results of cosmetic procedures. The system must be flexible enough to recognize a person despite natural changes, but at the same time not lose accuracy in protecting against fraudsters.

Systematic errors and demographic bias in face recognition algorithms

Algorithmic bias is a systemic problem in which recognition accuracy differs substantially for different demographic groups.

A typical situation: an algorithm trained predominantly on European faces shows an FRR of up to 35% for users from Africa or Asia. Women on average experience 20% more recognition problems than men. People over 60 encounter verification errors 5 times more often than a young audience.

For international platforms, this creates barriers to entering new markets. The inability to ensure an equal level of service for all users leads to discriminatory practices, lawsuits, and bans from regulators. In the US and the EU, requirements for the mandatory auditing of algorithms for bias are already in effect.

The sources of systematic errors are multiple: unbalanced training datasets, the architectural features of neural networks, cultural differences in the presentation of documents with photographs. Even technical factors such as the varying light sensitivity of cameras to different skin tones contribute to the unevenness of the results.

The solution requires a comprehensive approach: diversifying the training data, special fairness metrics, regular testing on representative samples, adaptive thresholds for different demographic segments. Ignoring the problem of bias inevitably limits the scalability of the business.

Implementation errors: wrong thresholds, camera quality, UX processes

Even an excellent algorithm can show unsatisfactory results with incorrect integration. Implementation errors often turn out to be the cause of 60-70% of all problems with biometric verification in production.

An incorrectly chosen decision threshold is the most common error. Using the vendor’s default values without adapting them to the specifics of the business leads either to excessive rejections of legitimate customers or to letting fraudsters through. The optimal threshold depends on the company’s risk appetite, the nature of the transactions, and the geography of the users. For money transfers it must be stricter than for viewing operation history.

The quality of image capture determines the upper limit of possible accuracy. A laptop’s frontal webcam with a resolution of 480p and a dynamic range of 6 stops is physically unable to capture enough biometric features. Insufficient or excessive lighting, glare on glasses, and motion blur reduce the quality to a level where even a human would struggle to recognize the face in the image.

UX processes directly affect the quality of the input data. The absence of clear instructions leads to users photographing themselves in unsuitable conditions. Overly strict requirements for face positioning cause frustration and refusals. The absence of feedback about the reasons for failure leaves the user confused.

The technical aspects of integration create additional sources of errors. Incorrect color conversion of images, excessive JPEG compression, incorrect handling of EXIF orientation metadata — each detail can critically affect the result. The absence of real-time quality monitoring makes it impossible to promptly identify and eliminate problems.

A comprehensive approach to implementation includes: pilot testing on real users, a gradual rollout with the possibility of a rollback, A/B testing of various configurations, continuous monitoring of quality metrics and customer feedback. Only systematic work on all the components of the integration ensures the stated accuracy figures of the algorithms in production conditions.

Factors affecting the accuracy of biometrics in KYC

The accuracy of biometric identification in KYC processes is determined by many variables that act at different stages: from the moment of capturing the biometric data to its processing by the algorithms. Even the most advanced neural network models can show substantially different results depending on the operating conditions and the technical implementation. Understanding these factors is critically important for configuring the system for specific business tasks and achieving the optimal balance between security and the user experience.

Image quality: lighting, angle, resolution, masks and accessories

The quality of the input data remains a fundamental limitation for any face recognition algorithm. Modern neural network architectures demonstrate impressive robustness to variations, but the physical limitations of the image still directly affect the probability of errors of the first and second kind.

Lighting creates the most substantial distortions of biometric features. Side light forms harsh shadows, which the algorithm may interpret as elements of facial geometry. Backlight makes key anthropometric points invisible. With insufficient illumination (less than 200 lux), the camera’s noise reduction activates, blurring the texture features of the skin and the boundaries of the facial features. According to NIST research, the transition from studio lighting to typical home conditions increases the FRR by 3-7% at a fixed security threshold.

The shooting angle determines the visibility of the biometric markers. A deviation from the frontal position of 15-20 degrees horizontally reduces the number of features available for analysis by 20-30%. Vertical deviations are even more critical: tilting the head down by 30 degrees makes the eye and nose-bridge areas — zones with the maximum concentration of unique biometric points — unavailable for analysis.

Image resolution determines the level of detail of the biometric features. The minimum inter-pupillary distance for reliable identification is 60-80 pixels. At lower values, algorithms lose the ability to analyze small details — the shape of the nostrils, skin texture, the microgeometry of the lips. This leads to an increase in both FAR (due to the loss of distinguishing ability) and FRR (due to the impossibility of finding a sufficient number of matching features).

Masks and accessories represent a separate category of complexity. Medical masks cover up to 60% of the face area, excluding the lower part of the nose, the lips, and the chin from analysis. Algorithms trained on data before 2020 showed a drop in accuracy of 20-50% when working with masks. Modern models, fine-tuned on mask datasets, have restored accuracy, but the FRR still remains 5-10% higher compared with uncovered faces. Sunglasses create similar problems, especially wide-frame models covering the eyebrows — a key area for biometric identification.

The capture channel: webcam, smartphone, photo vs video in the KYC process

The device and method of capturing biometric data form the basic accuracy limitations of the system. Each channel introduces specific distortions and determines the available verification methods.

Laptop webcams remain the most problematic channel for KYC. A typical resolution of 720p, a narrow dynamic range, and aggressive software image processing lead to the loss of up to 40% of biometrically significant information. A fixed focal length and the absence of optical stabilization make the distance to the camera critical: a deviation from the optimal 50-70 cm leads to blur or geometric distortions.

Smartphones provide substantially better capabilities. The frontal cameras of modern devices have a resolution of 8 megapixels or more, autofocus, and software distortion correction. The presence of a gyroscope makes it possible to control the orientation of the device and prompt the user for the optimal position. Computational photography (HDR, Portrait mode) improves quality in difficult lighting conditions, although it sometimes creates artifacts that complicate biometric analysis.

The choice between a static photo and a video stream determines the architecture of the entire verification system. A single photograph allows the user to choose the best frame, which reduces the FRR by 15-20%. However, this creates a vulnerability to attacks using someone else’s photographs. A video stream provides multiple frames for analysis, which makes it possible to choose the best one or average the results across several images. Processing a 3-5 second video reduces random recognition errors by 30-40% compared with a single shot, but requires a stable internet connection and greater computing power.

A hybrid approach — capturing a short series of photographs — combines the advantages of both methods. Analyzing 3-5 frames taken at an interval of 200-300 ms makes it possible to filter out blinks, micro-movements, and temporary occlusions, increasing the reliability of verification without substantially complicating the user experience.

The impact of the Liveness check on accuracy and face recognition errors

Technologies for confirming the presence of a live person (Liveness Detection) have become a mandatory component of modern KYC systems, but their integration creates additional challenges for the accuracy of biometric verification.

A passive Liveness check analyzes a single image or video stream without additional actions from the user. The algorithms look for signs of forgery: the absence of skin micro-textures, unnatural glints in the eyes, print or screen artifacts. These checks practically do not affect the basic recognition accuracy but can generate additional false rejections. Glare from glasses, unusual lighting, or low camera quality can be mistakenly interpreted as signs of an attack, increasing the system’s overall FRR by 2-5%.

Active Liveness requires the performance of random actions: head turns, blinking, smiling. These movements guarantee the three-dimensionality of the object and its real presence, but create additional complications for biometric comparison. A head turn of 30-45 degrees, necessary for a reliable Liveness check, reduces the number of visible biometric features. Facial movements deform the geometry of the face: a smile changes the shape of the eyes and cheeks, which can lead to a false rejection at strict verification thresholds.

Combined systems using several Liveness check methods simultaneously demonstrate the best protection against attacks but accumulate the errors of each method. A typical implementation with a passive texture check and an active movement check increases the overall FRR by 5-8% while maintaining the same level of FAR. This requires careful balancing of the thresholds: overly strict Liveness settings lead to rejections of legitimate users, overly soft ones open up opportunities for attacks.

The temporal synchronization between the Liveness check and the biometric comparison is critical for accuracy. The optimal strategy is to capture the reference image for biometrics at the moment of a neutral facial expression between active Liveness actions. This minimizes the impact of deformations and ensures the best conditions for comparison with the reference.

Aging of appearance and updating reference biometric templates

The change in a person’s appearance over time represents a fundamental problem for the long-term operation of biometric systems. Unlike static authentication methods, the biometric features of the face are subject to gradual transformation.

Natural aging causes predictable changes in biometric characteristics. Over 5 years, there is a decrease in skin elasticity, a change in facial contours, and the appearance of wrinkles. These changes increase the distance between the current image and the saved template in the biometric feature space. Studies show an FRR growth of 2-3% annually when using an unchanged reference. For people over 50, this degradation accelerates to 4-5% per year due to more pronounced age-related changes.

Short-term changes in appearance create additional variability. A weight change of 10-15% alters the geometry of the cheeks and chin. Growing or shaving a beard can increase the biometric distance by 15-20%. Cosmetic procedures, a suntan, temporary swelling — all of this affects the accuracy of comparison with a reference created in other conditions.

Template update strategies must balance between adapting to changes and protecting against identity substitution. Automatic updating after each successful authentication (template update) allows the system to adapt to gradual changes. But this creates the risk of template “drift” with systematic errors or targeted attacks. The accumulation of small changes can lead to a situation where the final template differs substantially from the original.

A hybrid approach with the storage of several template versions shows the best results. The system saves the original reference created at registration and 3-5 updated versions from recent months. During verification, comparison is carried out with all versions, and for authentication a match with one of them is enough. This reduces the FRR by 40-50% compared with a single static template, with minimal impact on FAR.

Periodic re-registration remains the most reliable method of maintaining accuracy. Requesting an update of biometric data every 2-3 years makes it possible to obtain a quality reference in controlled conditions and confirm the user’s identity by additional methods. For critical applications, this period can be reduced to 6-12 months, especially for age groups with rapid changes in appearance (teenagers, the elderly).

How to set accuracy requirements and manage face recognition errors in KYC

A properly built system for managing the accuracy of biometric algorithms determines the success of the entire KYC process. It is not enough to simply choose an algorithm with high figures — it is necessary to correctly configure it for the specific business requirements, regulatory constraints, and the features of your audience. Let’s consider a systematic approach to defining requirements and managing accuracy metrics in a production environment.

Defining acceptable FAR/FRR levels taking risk and regulators into account

Setting target FAR and FRR figures begins with an assessment of your business’s risk profile. For payment systems and crypto exchanges, FAR is critical — here it is unacceptable to let a fraudster through under the guise of a legitimate user. A typical requirement for high-risk operations: FAR ≤ 0.01% with an FRR of no more than 2%. This means no more than one false match per 10,000 verification attempts.

Financial regulators often dictate minimum requirements for biometric verification. The European PSD2 SCA standard requires a FAR of no more than 1/1000 for payment operations over EUR 30. The Bank of Russia’s requirements for remote identification set the FAR bar at 0.001% for higher-risk operations. At the same time, regulators rarely regulate FRR, leaving the balance between security and convenience to the business’s discretion.

For different types of operations within a single service, it is advisable to set differentiated requirements. Initial registration can work with a FAR of 0.1% and an FRR of 1%, ensuring smooth onboarding. Money transfers require tightening to a FAR of 0.001%. Logging into a personal account to view the balance allows softer settings: a FAR of 1% with a minimal FRR of 0.5%.

Industry specifics also affect the choice of thresholds. Online casinos and betting platforms focus on preventing multi-accounting, setting strict requirements of FAR < 0.0001%. Marketplaces and streaming services can allow a FAR of up to 1%, prioritizing a low FRR for maximum conversion of new users.

Threshold tuning and A/B tests for the “security versus conversion” balance

The optimal similarity threshold for making a decision about a face match rarely coincides with the algorithm’s default settings. The starting point for tuning is determined through an analysis of the similarity distribution on your real data. Collect a minimum of 10,000 “selfie-document” pairs from legitimate users and 1,000 attempts of known fraud attacks. Building a distribution histogram will show the zones of overlap between genuine and impostor attempts.

A/B testing of thresholds is conducted in stages with control of key metrics. Divide the traffic into segments of 10% and sequentially test thresholds with a change step of 0.05 on the similarity scale. Track not only the direct FAR/FRR metrics but also derived indicators: the percentage of successful registrations, the number of repeat attempts, the percentage of support requests with verification problems.

It is critically important to segment the results by demographic group and geography. Asian users may show a systematically higher FRR at the same thresholds as Europeans. The 60+ age group often requires lowering the threshold by 0.1-0.15 points to compensate for age-related changes in appearance and difficulties with selfies.

Dynamic threshold tuning based on the transaction context increases the effectiveness of the system. A new user from a low-risk country, registering with a verified email, is a reason to apply a softer threshold. An attempt to log in from a new device from another country via a VPN is a signal to tighten the requirements by 0.2-0.3 points on the similarity scale.

Monitoring the conversion of the verification funnel reveals the points of user churn. If more than 15% of users do not pass biometrics on the first attempt, the threshold is too high. If the success rate exceeds 98%, the system may be letting fraud attempts through. The optimal success rate of legitimate users on the first attempt is in the range of 92-95%.

Monitoring the algorithm’s quality in production and the regular re-evaluation of accuracy

The basic set of metrics for daily control includes: the number of successful/unsuccessful verifications, the average number of attempts per user, the distribution of similarity scores, the percentage of manual checks, the request processing time. Weekly reports should contain an analysis of FAR/FRR trends, segmentation by country and device, and the top reasons for rejections.

A quarterly accuracy audit is conducted on a control sample with manual labeling of the results. Randomly select 5,000 passed and 1,000 rejected verifications. The security team manually checks the correctness of the decisions, identifying the actual error figures. A discrepancy with the expected metrics of more than 20% is a reason for an unscheduled reconfiguration of the system.

Feedback from the support service and fraud analysts is critical for identifying systematic problems. Maintain a register of confirmed cases of false positives and misses. Analyze patterns: certain phone models, lighting conditions, ethnic groups. The accumulation of 50+ similar errors is a trigger for the targeted fine-tuning of the algorithm.

Retraining the model on the accumulated data is carried out once every six months. During this time, a sufficient volume of new examples accumulates, reflecting the current usage patterns. Include complex cases in the training sample: twins, age-related changes, deception attempts. Validation on a held-out sample must show an improvement in the metrics of at least 5-10%.

Technical debt accumulates with prolonged operation without updates. Once a year, conduct benchmarking against new versions of algorithms and competing solutions. The appearance on the market of models with metrics 30% better or more is a signal to consider migration. The transition plan must include the parallel operation of the old and new systems for at least a month to compare results on real traffic.

Documenting all setting changes and their impact on the metrics forms a knowledge base for the fast response to incidents. Each threshold change is recorded with an indication of the reason, the expected effect, and the actual result a week later. The change history helps avoid repeating unsuccessful experiments and speeds up decision-making in critical situations.