Bypassing Facial Recognition: Fraud Methods and Defenses

Ever since facial recognition was introduced, fraudsters have been looking for ways to bypass it – from simple paper masks to sophisticated deepfake technology. 

Sumsub's 2025–2026 Identity Fraud Report found that deepfake fraud attempts on its platform nearly doubled year over year in the UK, while the Maldives recorded a 2,100% increase. Synthetic media is rapidly becoming a weapon of choice for fraudsters targeting financial institutions.

The misuse of deepfake technology extends far beyond financial scams. In 2025, the FBI warned that North Korean IT workers were using false identities to obtain work at US companies to raise funds for the sanctioned regime. Growing evidence shows these fraudsters use real-time deepfake technology to animate stolen, purchased, or synthetic identities with AI-generated video and audio during remote job interviews and onboarding.

This article explores how fraudsters bypass facial recognition systems and what companies can do to verify that the person on screen is real, present, and legitimately linked to the identity being verified.

How facial recognition technology works

Facial recognition technology analyzes an individual’s facial features. It maps key biometric markers in a selfie or short video, such as the distance between the eyes or face shape, and compares them with a stored facial template. This face-matching process uses biometric data to assess whether the images belong to the same person.

Face verification vs. face recognition

Face recognition and face verification are often used interchangeably, but they work differently.

Face recognition uses a one-to-many comparison, checking a face against multiple identities to determine who the person may be. 

In contrast, facial verification uses a one-to-one comparison, matching a real-time image against a specific record, such as the face on an ID document. This comparison supports identity verification during the know your customer (KYC) process.

How safe is facial recognition for KYC?

Face biometrics help protect businesses and the public from deception, both during onboarding and in biometric authentication at later logins. Without anti-fraud mechanisms such as liveness detection, however, some systems can be fooled by spoofed images, pre-recorded videos, or AI-generated synthetic faces. That is why strong verification systems don’t rely on face matching alone.

In addition to biometric verification, comprehensive identity verification systems can check whether the person is physically present, whether the image or video has been manipulated, whether the device looks suspicious, and whether the identity data matches other risk signals.

This, along with privacy obligations, makes it critical for businesses to understand how their biometric facial recognition systems work and what their vulnerabilities are. 

Common face recognition use cases

Face matching supports identity verification in onboarding, account access, AML/KYC checks, and fraud prevention across heavily regulated industries such as banking, fintech, crypto, and iGaming.

As adoption has grown, so has the incentive for criminals to bypass these controls. Rather than attacking the underlying biometric algorithms, many fraudsters focus on deceiving the verification process itself – using fake faces, manipulated media, or compromised devices to appear legitimate.

Broadly speaking, these attacks fall into two categories: 

1. Presentation attacks, which attempt to fool the camera with fraudulent visual inputs
2. System-level attacks, which target the way biometric data is captured or transmitted

Presentation attacks

Presentation attacks: spoofing facial biometrics

Presentation attacks, also known as face spoofing, happen when fraudsters show the camera fake facial evidence to bypass biometric authentication. Instead of attacking the backend infrastructure, they try to trick the verification technology with something that looks like a real person’s face. Anti-spoofing controls are built to catch these attempts, which usually take one of the forms below.

Photo attacks with stolen selfies

Fraudsters can pull almost anyone’s photo from social media and use it to commit identity theft or other biometric fraud through face spoofing. If facial biometric technology doesn't analyze image characteristics like depth, movement, texture, or other signs of liveness, fraudsters can sometimes use social media images to bypass basic verification systems.

Similar methods can be used to access people’s bank accounts. For example, in 2023, a 34-year-old thief in Brazil accessed several accounts and applied for loans by placing customer photos over a dummy.

Replay attacks with pre-recorded video

Some systems ask users to blink, smile, turn their head, or follow on-screen instructions in a process commonly called active liveness detection. This can make presentation attacks harder for bad actors to carry out, but it does not automatically stop fraud. Fraudsters may use pre-recorded videos of a real person or manipulate existing footage to imitate the required actions to complete verification.

A deepfake video can copy someone’s face, reproduce expressions, or make it appear as if the person is completing a live verification check. If the system cannot tell whether the video is being captured live through a trusted camera, it may mistake replayed or manipulated footage for a genuine user.

Mask attacks: from paper to silicone

Masks fooled humans long before they fooled machines.

Fraudsters may also use hyperrealistic masks to impersonate someone else, hide their identity, or confuse a basic facial recognition system.

From 2015 to 2017, fraudsters used a custom silicone mask and a staged office to impersonate then-French Defense Minister Jean-Yves Le Drian over Skype, defrauding heads of state, wealthy individuals, and charities under the pretext of needing funds to pay ransoms for French hostages held by Islamist groups in the Middle East. Court filings and major reporting place victims' losses at approximately €55–80 million.

In China, the criminal use of lifelike silicone masks has long raised concerns about identity theft. Hyperrealistic masks, readily available on local e-marketplaces, are being marketed with claims they can fool facial recognition systems and let wearers impersonate others in high-security or workplace environments.

In December 2025, Chinese media reported that staff at a neighborhood committee in Wenzhou had been caught using printed paper masks of colleagues to bypass a facial recognition clock-in system.

Yet strong biometric systems do not rely on appearance alone. Advanced liveness and anti-spoofing checks can analyze whether the system is seeing real skin, natural facial movement, depth, blood-flow cues, and other signals that are difficult for a mask to reproduce convincingly, depending on the detection method and mask quality. For this reason, fraudsters are turning to AI-driven tools for more sophisticated deception.

AI-powered presentation attacks

Generative AI has lowered the cost and technical barriers to producing deepfakes and other fraudulent identity materials for presentation attacks. More sophisticated real-time attacks still require greater expertise and tools.

How deepfakes bypass biometric checks

One common attack is a face swap: the fraudster maps a victim’s face onto a video that mimics liveness cues such as blinking and nodding.

In May 2025, Vietnamese authorities dismantled a 14-person criminal ring that allegedly laundered VND 1 trillion (about US$38.4 million) by deploying AI-generated face biometrics to bypass facial recognition systems at banks.

In June 2026, an Amsterdam court sentenced a man to two and a half years in prison for opening 47 ABN AMRO accounts in other people’s names. He passed the bank’s app-based ID and selfie checks with deepfake-manipulated photos, and some of the accounts were later used in bank helpdesk fraud.

Effective deepfake detection can spot these attacks by analyzing artifacts in the submitted image or video.

Synthetic faces and synthetic identity fraud

Fraudsters can also use AI to create faces that do not belong to a real person. These synthetic faces may be paired with fake documents, stolen personal details, or fabricated account histories to create a convincing false identity.

This is closely linked to synthetic identity fraud, where criminals combine real and fake information to build an identity that can pass onboarding checks. For example, a fraudster may use a real national ID number with a fake name, a manipulated document, and an AI-generated face. If the system checks each element in isolation, the identity may appear plausible.

During identity verification, a synthetic face may not match any known victim because the person does not really exist.

Multimodal deepfake attacks

Deepfake fraud is becoming more multimodal: criminals combine several types of synthetic or manipulated content in one attack, which makes identity verification harder.

A fraudster may use an AI-generated face for onboarding, a deepfake video for liveness, a forged ID for KYC, and a cloned voice for customer support calls. Together, these deepfakes and forgeries reinforce one another, so the identity holds up at every checkpoint.

Bypassing facial age estimation checks

Unlike facial verification, facial age estimation assesses whether a person is likely to meet an age threshold without confirming their identity. Attempts to bypass age verification may involve face spoofing using an adult’s photo, replayed video, or manipulated media.

Defenses against face spoofing should combine age estimation with liveness checks to help establish that a live person is present. A challenge age above the legal threshold can route uncertain results to another verification method.

System-level attacks on biometric pipelines

Not all biometric attacks involve showing a fraudulent face to the camera. Some fraudsters try to bypass facial recognition by attacking the verification flow itself. In these system-level attacks, criminals manipulate how biometric data is captured, transmitted, or submitted to the platform.

An injection attack can replace the expected camera input, while a relay attack can redirect a genuine interaction into a fraudulent session.

Camera injection attacks 

Camera injection attacks attempt to bypass liveness detection by feeding fake biometric data directly into a verification system, like an app or API, rather than using the camera. Fraudsters can hijack a camera stream with synthetic or pre-recorded footage using a virtual camera, tamper with the app using emulators or rooted devices, or manipulate an API by sending generated or replayed biometric data.

These attacks often succeed when systems lack strong endpoint security or client integrity checks. To defend against them, advanced liveness solutions use secure enclaves, anti-tampering SDKs, and encrypted pipelines to verify the integrity of both the device and the data.

Relay attacks and session hijacking

Relay attacks occur when a real person performs the required liveness actions, but the interaction is forwarded remotely from a different location to trick the biometric authentication system. For instance, a fraudster might socially engineer a victim into a video call and secretly route that video to the liveness check. These attacks are particularly effective against systems that don’t validate the camera input's origin or track geolocation.

Session hijacking involves taking over an existing session. Both threats can undermine identity verification even after a real person completes a check. Robust defenses against relay fraud include session validation, IP and device fingerprinting, geofencing, and behavioral or contextual analysis to confirm whether the user is both live and in the expected location.

Liveness systems are only as reliable as their weakest link: typically, the user’s device, the transmission layer, or the backend server. Each layer must be secured to prevent spoofing, relay attacks, or tampering.

Telegram KYC bypass kits and fraud-as-a-service

Illicit marketplaces turn biometric attacks into fraud-as-a-service. An April 2026 MIT Technology Review investigation identified multiple public Telegram channels and groups advertising KYC bypass kits and stolen biometric data. Sellers promoted tools for camera injection using virtual cameras to replace live input with other images or video.

Some underground services combine forged documents, AI-generated facial videos, and stolen identity data into full KYC bypass solutions. These deepfake KYC services may target the submitted media, the capture process, or both. 

Real-world deepfake and biometric fraud cases

Recent cases show how deepfake fraud and identity theft can undermine KYC checks, with biometric fraud targeting both submitted images and live video verification.

In a June 2026 judgment, an Amsterdam court found that 47 ABN AMRO bank accounts had been opened fraudulently using manipulated selfies and identity documents. Posing as a landlord, the defendant obtained prospective tenants’ identity documents. He then used deepfake methods to blend his facial features with theirs and open the accounts.

In August 2026, a Spanish National Police investigation described a suspect who allegedly made 38 attempts to obtain electronic certificates using more than 30 stolen identities. He allegedly used forged IDs and real-time facial manipulation to try to make his face match the document presented during video identification.

During one verification call, the digital mask briefly disappeared, exposing his real face to the certificate provider’s security staff and helping investigators identify him.

Biometric fraud is no longer limited to fake selfies. Fraudsters combine altered documents, synthetic media, stolen data, and weak verification flows to make false identities appear real.

Suggested challenge: Could you have stopped the biggest financial crimes in history? Step into Sumsub’s AI Time Machine. 

Biometric data security and compliance

Facial recognition strengthens identity checks, but it also involves sensitive biometric information. This makes biometric data security essential for regulatory compliance.

In many jurisdictions, biometric data receives special protection when it is used to uniquely identify a person. Under the EU GDPR and UK GDPR, biometric data processed to uniquely identify a natural person is treated as special-category data under Article 9. Organizations must identify both an Article 6 lawful basis and an applicable Article 9(2) condition (for example, explicit consent or substantial public interest under domestic law), and apply appropriate technical and organizational safeguards before processing it.

Security and privacy controls should be built into the verification flow from the start. Key safeguards include:

• collecting only the data needed for the identity check
• clearly explaining how biometric data will be used
• protecting biometric templates and verification results with strong encryption
• limiting access to biometric records
• setting clear retention periods
• monitoring for unauthorized access, tampering, and misuse

Presentation attack detection should also be tested against recognized standards. ISO/IEC 30107-3:2023 is the current international standard for testing and reporting the performance of biometric presentation attack detection mechanisms.

Strong biometric data security means protecting the user’s biometric information throughout its lifecycle, from capture and processing to storage, audit, deletion, and fraud monitoring.

EU AI Act deepfake rules from August 2026

The EU AI Act’s Article 50 transparency rules apply from August 2, 2026. Deployers must disclose when image, audio, or video content constitutes a deepfake. Providers of systems that generate synthetic content must mark outputs in a machine-readable format so they can be detected as AI-generated or manipulated.

For systems placed on the market before August 2, 2026, the provider marking obligation applies from December 2, 2026, according to the European Commission. Certain authorized law-enforcement uses are exempt, and artistic or satirical works have tailored disclosure requirements.

Rules on deepfakes complement GDPR protections for biometric data. 

Injection attack detection standards

ISO 30107-3, formally ISO/IEC 30107-3:2023, covers testing and reporting for presentation attack detection. Attacks that bypass the capture device, such as camera injection attacks, fall outside its scope.

CEN/TS 18099:2024 is a separate testing framework for injection attack detection, addressing biometric data injection attacks, mitigation, and evaluation planning. ISO/IEC 25456, an international project on standards for biometric data injection attack detection, remains under development as of October 2026.

How liveness detection stops spoofing

Liveness detection is the core of facial anti-spoofing. It checks whether the face shown during verification belongs to a real person who is physically present, rather than to a photo, mask, replayed video, deepfake, or manipulated camera feed. Face matching only confirms that two faces look similar; it does not prove the user is live.

Businesses use face liveness detection to spot spoofing attempts during onboarding, login, account recovery, and other high-risk actions. For a full breakdown of how it works, see our dedicated Liveness guide and Liveness product page.

Liveness alone can’t stop sophisticated fraud, though, as the signals below show.

Liveness signals biometric systems analyze

Different liveness detection systems look for different signals to help protect biometric authentication against face spoofing. The strongest ones combine several at once. Common categories include:

• Micro-movements: Real faces make small involuntary movements that are very hard to fake, such as microsaccades (tiny eye movements), pupil responses to changes in light, and slight asymmetries in natural blinking.
• Skin and texture: The system examines pore structure, subsurface light scattering, and how skin reflects light at different angles. Masks, screens, and printed photos reflect light very differently from human skin.
• Depth and 3D structure: The check confirms that the face has real three-dimensional geometry rather than the flat profile of a photo or screen. Some systems use structured light or stereo cameras, while others infer depth from how shadows and highlights shift during a session.
• Physiological signals: Remote photoplethysmography (rPPG) detects the tiny changes in skin color caused by each heartbeat. Static images and most deepfakes don’t reproduce this signal correctly.
• Camera and environmental consistency: The system checks whether lighting, focus behavior, and image noise match a real camera filming a real scene, rather than a video played back to the camera or injected into the session.
• Frequency-domain artifacts: Frequency analysis shows whether the image carries the compression patterns and sensor signatures of a genuine camera feed, or traces of AI generation or video re-encoding.

The strongest liveness systems don't rely on any single signal. Each technique has a known failure mode: depth checks can sometimes be fooled by high-quality 3D masks, rPPG can be approximated in advanced deepfakes, and texture analysis can struggle in poor lighting. Layering multiple signals significantly increases the cost of an attack, because the fraudster must defeat all of them simultaneously in a single session.

Layered defenses beyond liveness detection

Effective fraud prevention relies on a multi-layered approach that combines liveness detection, deepfake detection, device and session integrity checks, document verification, behavioral and risk signals, and, where appropriate, human review.

Device fingerprinting helps identify suspicious devices and links between apparently separate accounts. Deepfake detection assesses manipulated media, while injection attack detection checks for interference with the capture process. Checks should continue beyond onboarding, with additional verification for suspicious logins, account recovery, or high-risk transactions.

Suggested read: Liveness Detection: A Complete Guide for Fraud Prevention and Compliance

How to choose a spoof-resistant liveness solution

A strong liveness solution should catch common spoofing methods without making verification difficult for genuine users.

Presentation attack detection (PAD) covers anything held up to the camera, such as printed photos, screen replays, and masks. When comparing vendors, ask for both error rates: how often attacks get through (APCER) and how often genuine users are wrongly rejected (BPCER). A system that blocks every attack but turns away one in ten real customers will cost you conversions.

Deepfake detection covers AI-generated and manipulated faces, including face swaps and fully synthetic identities. Because generative tools change quickly, ask how often detection models are retrained and whether they analyze video as well as still images.

The two overlap. A deepfake played on a phone screen in front of the camera is a presentation attack; the same deepfake streamed through a virtual camera is an injection attack. Neither PAD nor deepfake detection stops the second on its own, so the solution should also verify that the camera feed hasn’t been tampered with.

Active vs passive liveness detection

Active liveness detection asks users to perform an action, such as blinking, smiling, or turning their head. Passive liveness detection works in the background while the user takes a selfie or short video. Active liveness checks may add friction, while passive checks may be more user-friendly.

Testing liveness against real attacks

A liveness solution should be tested against realistic attack types, including printed photos, replayed videos, masks, and deepfakes presented to the camera. Presentation attack detection testing, including checks aligned with ISO 30107-3, can help businesses assess how well a system detects spoofing attempts. Camera injection requires separate testing because it falls outside the scope of ISO 30107-3.

Suggested challenge: Is your memory good enough to beat the Beast? 

Bypassing facial recognition FAQ

• How does facial recognition work?

  Facial recognition captures a face image or video, detects the face, maps key facial features, creates a biometric template, and compares that template with a trusted image or record. In onboarding, this usually means comparing an image or video uploaded by the user with the photo on their identity document.

• Can facial recognition be fooled?

  Yes, facial recognition can be fooled by photos, replayed videos, masks, or deepfakes when appropriate safeguards are missing. Strong systems use deepfake detection, liveness detection, device checks, and other fraud signals to spot manipulated videos or synthetic faces.

• Is facial recognition safe?

  Yes, with appropriate safeguards. Facial recognition can support identity verification more robustly when combined with liveness detection, document verification, encryption, secure data handling, and ongoing fraud monitoring. Businesses should also limit biometric data collection, access, and retention.

• What is liveness detection?

  Liveness detection verifies that the face in a selfie or video belongs to a real person physically present during the verification session. It may analyze signals such as movement, depth, skin texture, lighting, image consistency, and camera behavior to detect photos, masks, replayed videos, deepfakes, or injected media.

• What is biometric verification?

  Biometric verification is the process of confirming a person’s identity using a unique physical or behavioral trait, such as their face or fingerprint. In facial verification, the system compares a live selfie or video with a trusted image, such as the photo on an identity document.