Deepfake Scams in Southeast Asia: How Banks Can Detect Fake Faces

How deepfakes get past bank face checks in Southeast Asia, what MAS, BNM, BOT, BSP, OJK, and SBV ask banks to check, and what each check catches.
A deepfake is an image, video, or audio clip created or altered with AI. In Southeast Asia, deepfakes have been used against the face checks banks run in electronic Know Your Customer (eKYC) onboarding and when a customer approves a high-value transfer. Liveness detection looks at what is shown to the camera, and injection controls look at feeds that bypass it; no single check covers both.
How deepfake attacks on eKYC and bank face checks work
The Monetary Authority of Singapore’s (MAS) information paper of September 2025 says deepfakes “have the potential to defeat biometric authentication solutions used by FIs” (financial institutions), at onboarding and at “subsequent login to financial services using biometric authentication”. The cases in the next section add a third moment: approving a transfer above a threshold.
Presentation attacks: a fake face shown to the camera
In a presentation attack, the fake is held up to the camera: a printed photo, a replayed video, a mask, or a deepfake played on another screen. The frames come from a real camera, and the question is whether the face is live. The ISO/IEC 30107 series on presentation attack detection (PAD) covers attacks made at the capture device.
Injection attacks: a fake feed in place of the camera
In an injection attack, manipulated or pre-recorded media enters the verification flow without passing through the physical camera. MAS names three routes at a high level: virtual camera software, scripts injected into a browser, and man-in-the-middle interception. Injection attacks fall outside ISO/IEC 30107, so a presentation-attack test alone does not cover them.
Cost and effort
No regulator, police, or UN source cited here gives a cost figure for these attacks. Vendor price figures for deepfake tools exist, but they are marketing claims and are not used here. A footnote to the Bank of Thailand’s (BOT) guideline grades fake biometrics by effort, from face data taken from social media (low) and video used to mimic liveness detection (medium) to a 3D mask (high).
The United Nations Office on Drugs and Crime (UNODC) identified more than ten deepfake software providers targeting fraud groups in Southeast Asia, mainly through Telegram, offering deepfakes as a service; it gives no price. It also reported “more than a 600 per cent increase in mentions of deepfake-related content targeting criminal groups across monitored platforms between February and July 2024”. These are mentions on monitored Telegram marketplaces, not incidents.
Tools, settings, and step-by-step methods are left out here.
What has happened in Southeast Asia
- Indonesia, August 2024. MAS says an Indonesian financial institution reported that deepfake photos defeated its digital Know Your Customer (KYC) process for loan applications; MAS cites Group-IB’s analysis of the case.
- Vietnam and Thailand, February 2024. MAS, also citing Group-IB, says media stolen from victims’ phones by malware were used to build deepfakes in “attempts to trick the facial biometric authentication of banks”. The paper does not say whether those attempts succeeded.
- Vietnam, May 2025. Thái Bình police described a gambling network that used AI-made clips of account holders’ faces to pass bank biometric scans.
- Vietnam, October 2026. Tuổi Trẻ reported on 8 October 2026 that the State Bank of Vietnam’s (SBV) payment director said deepfake and AI tools can get past some biometric checks, and that institutions had been asked to review, patch vulnerabilities, and replace old technology. No SBV text of it was found.
The Thái Bình statement gives three steps. Hired people opened bank accounts, and a person abroad ran the banking apps on their phones remotely. For transactions of VND 10 million or more, which the police said require a biometric scan, AI clips of each holder’s face passed the check. Police found the scheme through their own work; the statement does not say what the banks’ checks observed. This is mule-account automation: the deepfakes stood in for hired holders. Sources: the police statement reprinted by Tạp chí Thị trường Tài chính Tiền tệ on 1 June 2025, and VnExpress, 29 May 2025.
Group-IB’s own, vendor-reported account of the Indonesian case (4 December 2024) adds that the photos came through a virtual camera, past anti-emulation, application self-protection, face recognition, and liveness detection, and that its team found “over 1,100 deepfake fraud attempts” from “45 specific devices”.
Our inference from this account: the checks saw what looked like a live camera capture, and the pattern was visible at the level of devices shared across many applications, which a single face check does not see.
What regulators ask banks to check
Six regulators’ texts are cited here: MAS, Bank Negara Malaysia (BNM), the BOT, Bangko Sentral ng Pilipinas (BSP), Indonesia’s Otoritas Jasa Keuangan (OJK), and the SBV.
| Regulator and instrument | Genre and modality | What it says about face checks |
|---|---|---|
| MAS “Cyber Risks Associated with Deepfakes” | Information paper; “should”, “could” | Document checks, liveness detection, deepfake testing, injection controls |
| BNM e-KYC policy document | Standards (“S”, shall) and guidance (“G”, encouraged) | Three modules, mandatory for higher-risk accounts; liveness test minimums |
| BOT Guideline for Biometric Technology Adoption | Guideline; “should”; unofficial English translation | Liveness detection “when possible”; accuracy targets; attempt limits; failed-check protocol |
| BSP Circular No. 1213 | Circular; “must” | Device, app-integrity, and authentication controls; no deepfake or liveness mention |
| OJK POJK 8/2023 and POJK 21/2023 | Regulations; “must” | Biometric factor in remote onboarding; no deepfake or liveness wording |
| SBV Circular 50/2024, as amended by Circular 77/2025 | Binding circular (“phải”, must) | Certified PAD, ISO 30107 Level 2 or equivalent; app integrity; transfer biometrics |
Singapore: MAS’s information paper
MAS’s information paper groups mitigations into six areas, four of them in the table above. MAS noted that deepfake incidents after onboarding had not been reported in Singapore.
Malaysia: BNM’s e-KYC policy document
BNM’s policy document on electronic Know-Your-Customer (e-KYC), issued and effective on 15 April 2024, covers institutions including licensed banks, life insurers, money services businesses, and approved issuers of designated payment instruments. In general, an institution “may” use three modules: document verification, biometric matching, and/or liveness detection (G 8.10). The technology provider’s solution must be assessed by a credible external independent assessor (S 8.22).
For higher-risk products, the modules are mandatory (S 8.32–8.33, Appendix 4). The products are current accounts, savings accounts, and unrestricted investment accounts “with funds placement and withdrawal flexibilities as well as funds transfer features”. The rule covers “existing individual customers” who hold none of them and apply through eKYC (footnote 31). For them, an institution “shall” at minimum, among other measures, verify the customer against a government-issued ID biometrically, check that the ID is authentic, and use liveness detection.
Thailand: the BOT guideline
The BOT’s Guideline for Biometric Technology Adoption in Financial Services says providers should set procedures to detect impostors at registration, for example using “presentation attack detection or liveness detection technology, and other controls when possible”. For face comparison against trusted sources, it recommends a False Acceptance Ratio below 0.1% and a False Reject Ratio below 3% (footnote 3, citing the US National Institute of Standards and Technology’s (NIST) SP 800-63B and Fast Identity Online (FIDO)). Separately, a BOT press release (No. 10/2023, 9 March 2023) says a facial scan is needed, for example, to open a bank account through a banking app or make a digital transfer of more than 50,000 baht per transaction.
The Philippines and Indonesia: BSP and OJK
BSP Circular No. 1213, Series of 2025, implements Section 6 of the Anti-Financial Account Scamming Act (AFASA). Its safeguards “must include” restrictions on installing apps on rooted or jailbroken devices and emulators, device fingerprinting with anti-spoofing, and blocking unauthorized scripts and automation.
OJK rules require a biometric factor in fully remote onboarding (POJK 8/2023, Article 21(5)(c)) and, for commercial banks, in video and remote verification (POJK 21/2023, Article 5(7)).
Vietnam: Circular 50/2024 as amended
The SBV amended its online-banking security rule, Circular 50/2024/TT-NHNN, with Circular 77/2025/TT-NHNN, in force since 1 March 2026. The SBV named deepfakes and malware as the reason. It applies to credit institutions, foreign bank branches, payment intermediaries, and mobile-money providers, among others.
The presentation attack detection used for biometric matching must be certified to ISO 30107 Level 2 or equivalent. The certificate can come from a FIDO Alliance-recognized laboratory or from a certification body accredited under the International Accreditation Forum (IAF) multilateral arrangement. The rule applies whether the bank builds the check itself or buys it from a vendor.
For individuals, a transfer above VND 10 million, or one that takes the day’s total above VND 20 million, needs biometric matching on top of a one-time password (OTP), FIDO, or e-signature factor (Circular 50/2024 annexes, as replaced by Circular 77/2025). The replaced annexes apply from 1 July 2026 to providers that serve both individuals and organizations.
Deepfake detection layers: what each check catches and misses
| Layer | Catches | Misses |
|---|---|---|
| Document checks | Tampered IDs; AI content with a fingerprint or watermark | AI content without a fingerprint or watermark |
| Face matching | A selfie of a different person | The same synthetic face on the ID and the selfie |
| Liveness detection | Photos, replays, and some masks shown to the camera | Injected feeds; attack types passive tests handle poorly |
| Injection, device, and session signals | Feeds not from the camera; tampered apps; mismatched devices | Attacks that look like a normal device, as in Indonesia |
| Step-up and review | Cases an earlier layer has flagged | Attacks that no layer flags |
Document checks
MAS says FIs should use “robust methods” to verify ID documents, among them metadata analysis, image forensics, fingerprinting (“unique patterns unintentionally embedded by AI models”), and watermarking. It notes that “some AI tools may not inject fingerprints or watermarks”.
Face matching
Face matching, also called face verification, compares the selfie with the ID or registry photo. Our reading: a match shows only that two faces belong together, so a forged ID and a selfie carrying the same synthetic face can pass.
Liveness detection
Liveness detection, as ISO/IEC 30107-1 defines it, is analysis to determine whether a sample “is being captured from a living subject present at the point of capture”. MAS says FIs should implement it “to detect manipulated videos through motion analysis, texture analysis, thermal imaging, 3D Depth analysis and behavioural analysis”, and could prompt specific actions. The paper names these signals without explaining them.
The EU cybersecurity agency (ENISA) described what they measure in a 2022 report. Active checks ask the user to do something unpredictable, such as turn, blink, or follow a moving dot, and check that the response matches. Fast movements are hard for real-time deepfake software to render. Passive checks need no action. They look for involuntary signals, such as skin-color changes with the pulse and natural blinking, and for traces of synthesis, such as face-warping artifacts and resolution mismatches. Depth checks map the 3D geometry of the face. ISO/IEC 30107-1 counts involuntary reactions, such as “heart activity (pulse)”, and voluntary ones, such as “a biometric presentation in response to a directive cue”, as evidence of liveness.
NIST’s 2023 test of passive, software-based detectors found that they caught photo prints, replays, and some masks well, though “only a small percentage of algorithms” showed notable detection capability. Error rates on other attack types were “high across all algorithms tested”, and the report is “silent on … injection attacks”.
Injection, device and session signals
MAS says FIs should use injection detection software development kits (SDKs) to identify “suspicious patterns and anomalies during verification attempts”, and that runtime application self-protection (RASP) can also be used. In Vietnam, the mobile banking app must exit or stop, and tell the customer why, if it detects a debugger, an emulator, injected code (hooking), a repackaged app, or a rooted or jailbroken phone (Article 8(4), as amended). BNM lists device indicators an institution may use, such as IP address, geo-location, and device IDs (Appendix 4).
NV Global’s liveness product is deepfake-resistant and is also resistant to video stream injection.
Step-up and manual review
Among the texts cited here, two set out what happens after repeated failed checks. Vietnam’s circular requires (“phải”) that biometric confirmation be locked after the institution’s limit of consecutive failures, at most 10, and unlocked only at the customer’s request after the customer is checked (Circular 50/2024 Article 11(5)(d)). Thailand’s BOT guideline recommends a sequence (Principle 3(3), footnote 7): fewer than 10 failed attempts for presentation attack detection and fewer than 5 otherwise, then an exponentially growing delay, then refusing authentication and recommending another method.
For genuine customers, the BOT guideline asks providers to define a protocol for failed biometric verification, such as alternative verification procedures or alternative documents from trusted sources (Principle 3(4)). BNM’s e-KYC FAQ (question 4) says “false negative customers should not be immediately dismissed” and that “remedial actions should be considered where the customers can prove authenticity of their identification”.
Testing against standards
MAS says FIs should regularly test liveness detection against deepfake samples, across datasets. BNM’s Appendix 3 sets minimum scenarios for the provider’s liveness test, which “may be done in conformance to ISO/IEC 30107-3”: 2D masks, 3D masks, images altered with “ShallowFake” software, “DeepFake” tests “if possible”, and replay and man-in-the-middle tests. Vietnam is the only regulator cited here that names a certification level for presentation attack detection.
| Standard or report | What it covers | Status |
|---|---|---|
| ISO/IEC 30107-3:2023 | Testing and reporting of presentation attack detection | Published |
| CEN/TS 18099:2024 | Biometric data injection attack detection; presentation attacks out of scope | Technical Specification, valid from 16 December 2024 |
| ISO/IEC 25456 | Biometric data injection attack detection | Under development at ISO |
A checklist for fraud and onboarding teams
The steps below apply to any identity verification flow with a face check and name the source of each.
List camera, upload, browser, and app paths and the control that sees each (MAS).
Look for tampering and AI fingerprints or watermarks (MAS; BNM G 8.10, mandatory under Appendix 4 for higher-risk accounts).
Run them as distinct steps (BNM G 8.10 and Appendix 4; MAS).
Use injection detection, app integrity checks, and install restrictions (MAS; BSP Circular 1213; in Vietnam, Circular 50/2024 Article 8(4), as amended).
Use device indicators (BNM Appendix 4, optional) and device fingerprinting (BSP Circular 1213, mandatory). Flag many applications from one device: in the Indonesian case, Group-IB reports over 1,100 attempts from 45 devices (vendor-reported).
Cover at least BNM’s Appendix 3 scenarios, and test liveness against deepfake samples (MAS).
For any ISO/IEC 30107-3 claim, ask who tested, which attack types, and the attack presentation classification error rate (APCER) and bona fide presentation classification error rate (BPCER) for each. A rate from one lab is not necessarily reproducible by another.
Ask how injection was tested against CEN/TS 18099:2024 and, where BNM applies, for the external assessment (S 8.22). In Vietnam, ask for the ISO 30107 Level 2 (or equivalent) certificate and who issued it.
Follow the BOT sequence (Principle 3(3), footnote 7) and, in Vietnam, lock biometric confirmation after at most 10 consecutive failures (Circular 50/2024 Article 11(5)(d)).
Offer an alternative verification procedure (BOT Principle 3(4)); in Malaysia, do not dismiss false-negative customers immediately (BNM FAQ, question 4).
In Malaysia, adopt immediate mitigation measures when a weakness in the eKYC solution is found (S 8.29, “shall”). Heightened inspection sampling, back-testing, and a work plan are optional (BNM FAQ, question 15).
In Malaysia, sample at least 400 cases a month, or a 95% confidence level with a 3% margin, whichever is higher. Notify BNM within seven working days if the false acceptance rate exceeds 5% in any two months within four (Appendix 2).
BNM sets revalidation at least every three years or after a material change.
This article is general information, not legal advice. Sources were checked on 17 November 2026. Vietnamese texts are quoted in NV Global’s unofficial translation, and the BOT guideline is quoted from its unofficial English translation. Institutions should confirm how these texts apply to them with their regulator or legal counsel.
The deepfake attacks described in the sources cited here target three moments: onboarding, later biometric login, and the approval of transfers above a threshold. They enter either in front of the camera or in place of it. The region’s regulators describe the checks in different genres: MAS recommends, BNM sets standards and makes the core checks mandatory for higher-risk accounts, the BOT guides, the BSP requires device and authentication controls without naming deepfakes, and Vietnam’s SBV requires certified presentation attack detection.
Taken together, the texts suggest that each check, from the document to the device, covers a different gap. Where the cited cases succeeded, they passed checks already in place, so testing and review after a miss matter as much as the checks themselves.