Singapore is both a prime target for deepfake fraud and one of the most proactive jurisdictions in regulating it. Scam victims in Singapore lost over 913 million dollars in 2025, and the tactics now routinely include deepfake impersonation of executives, bank officers, and even national leaders whose likenesses have appeared in fraudulent investment ads. In response, the Monetary Authority of Singapore and other agencies have moved faster than most regulators anywhere, setting out concrete expectations for how financial institutions must defend against synthetic media.
This piece explains what Singapore's regulators now require and expect in relation to deepfake fraud: why the country is such a target, the measures MAS has laid out for financial institutions, the Shared Responsibility Framework that puts scam losses on institutions, the compliance stakes, and what it all means in practice. It is written for the banks, payment providers, and fintechs that operate under MAS supervision. One note: this area is moving quickly, and the authoritative source is always MAS's current guidance, so treat this as an orientation rather than legal advice.
The throughline is that Singapore has decided deepfake defense is not optional. Between binding frameworks and clearly stated supervisory expectations, financial institutions are now accountable for whether their systems can tell a real customer from a synthetic one.
- Singapore is a high-value target: scam victims lost over $913 million in 2025, and deepfakes now impersonate executives, bank staff, and even national leaders.
- MAS is among the world's most proactive regulators on this, and its September 2025 deepfake information paper sets out clear expectations for financial institutions.
- MAS expects FIs to verify document and content authenticity, use liveness that covers deepfakes, add endpoint protection against injection, simulate deepfake attacks, and integrate real-time deepfake detection.
- The Shared Responsibility Framework, effective December 2024, places phishing-scam losses on financial institutions and telcos that breach their anti-scam duties.
- FIs must run real-time fraud surveillance and hold rapid account-draining transactions for at least 24 hours, with 12-hour delays on higher-risk activities.
- A deepfake-opened account is not just fraud but a KYC and AML compliance failure under MAS Notice 626, raising the regulatory stakes.
- A real documented case saw a finance director nearly transfer millions after a deepfake video call impersonating the CEO and colleagues.
- Meeting these expectations means layered, deepfake-aware detection at onboarding, re-verification, high-value transactions, and voice channels.
Why Singapore Is a High-Value Target
Singapore's profile is close to the opposite of a high-volume emerging market, and that is exactly what makes it attractive to fraudsters: concentrated wealth, a dense financial-services sector, near-universal digital banking, and a role as a regional hub, all of which raise the payoff per successful attack. Global and regional fraud syndicates target it accordingly, and the losses are substantial, with scams costing victims more than 913 million dollars in 2025. The dominant 2026 tactics reported against Singapore institutions include deepfake impersonation of executives to authorize fraudulent payments, mule networks laundering the proceeds, and social engineering across messaging and payment rails.
The deepfake dimension is concrete, not hypothetical. Fraudsters have circulated manipulated videos of Singapore's own leaders endorsing bogus investment schemes, and executive impersonation has produced real losses. In one documented case, a company finance director was lured into a fake video conference featuring a deepfake of the company's chief executive and colleagues, and instructed to transfer corporate funds, moving over 499,000 dollars before a further request for 1.4 million dollars raised suspicion and the police anti-scam centre intervened. That single case captures why regulators treat this as urgent: the fake was convincing enough to move real money in a sophisticated corporate environment.
What MAS Now Expects of Financial Institutions
The centerpiece is a MAS information paper on the cyber risks of deepfakes, published in September 2025, which sets out the mitigating measures MAS expects financial institutions to adopt. It is not a vague call to awareness; it is a specific list. MAS recommends that institutions conduct comprehensive checks to detect tampering and verify the authenticity of identification documents and digital content during onboarding and verification, implement liveness detection in biometric authentication that specifically covers deepfakes, and implement endpoint-level protection to prevent injection attacks. It further recommends regular vulnerability assessments that simulate deepfake attacks to test the resilience of biometric systems, strong encryption of biometric data, and, notably, real-time deepfake detection integrated into communication channels and endpoints, alongside additional verification for high-risk transactions.
Two things stand out about that list. First, it names injection attacks and deepfake-aware liveness explicitly, which means MAS understands that ordinary liveness and a presentation-attack certificate are not sufficient. Second, it treats detection as something to be built into live channels and tested against simulated attacks, not bolted on once. These expectations sit on top of binding obligations: a deepfake that defeats onboarding does not just cause a fraud loss, it can constitute a know-your-customer and anti-money-laundering failure under MAS Notice 626, which turns a detection gap into a compliance breach.
Six measures every MAS-supervised institution is now expected to run
Table 1: The measures MAS recommends financial institutions take against deepfakes.
The Shared Responsibility Framework and Anti-Scam Duties
Where the deepfake paper sets expectations, the Shared Responsibility Framework sets binding duties with money attached. Announced by MAS and the Infocomm Media Development Authority in October 2024 and effective from 16 December 2024, the SRF allocates responsibility for losses from a defined scope of phishing scams among financial institutions, telcos, and consumers, and it requires payouts to affected victims where an institution breached its duties. For banks and major payment providers, those duties include real-time fraud surveillance aimed at the rapid draining of significant account balances, with a requirement to block or hold such transactions for at least 24 hours so customers can review them. Separately, since October 2025, banks have applied delays of 12 hours or more on higher-risk activities such as adding a new payee or raising a transfer limit, safeguards MAS credits with reducing impersonation-scam losses.
It is worth being precise about the SRF's scope, because it is often overstated. It covers phishing scams with a digital and territorial nexus, and it excludes malware-enabled scams and authorized-payment scams such as investment or romance scams where the victim intended the transfer. In other words, the SRF is one important lever, not a complete answer to deepfake fraud, much of which arrives through authorized-payment and impersonation routes it does not directly cover. That gap is part of why MAS pairs the binding SRF with the broader supervisory expectations of the deepfake paper, and why the Deputy Chairman of MAS has framed the response as one where scam prevention requires a multi-layered approach.
Table 2: Singapore's core regulatory levers against scam and deepfake fraud.
The Compliance Stakes
For an institution, the effect of this regulatory stack is that deepfake defense has moved from a security nicety to a supervised obligation with financial and reputational consequences. On the loss side, the SRF can require payouts to victims when anti-scam duties were not met, shifting cost onto the institution rather than the customer. On the compliance side, deepfake-enabled onboarding fraud, AI-generated faces that pass photo liveness, deepfake video replayed during verification, manipulated identity documents matched to fabricated faces, or an injected camera feed, produces fraudulent accounts that expose the institution to KYC, AML, and counter-terrorism-financing failures under MAS notices, with remediation costs and potential enforcement attached. And the attack surface is not only onboarding: Singapore's major banks run large voice channels for customer service and authentication, and voice cloning can now replicate a customer or executive from seconds of audio, extending the exposure to the call center.
The regional regulatory direction reinforces this. Across ASEAN, authorities have responded to deepfake and injection fraud with stricter eKYC, biometric mandates, and enhanced AML requirements, a clear shift toward higher-assurance digital identity. The consistent conclusion, in Singapore and among its neighbours, is that regulation alone cannot beat adversaries using AI at industrial scale, and that institutions need an AI-versus-AI posture combining liveness, injection detection, biometric forensics, behavioral analytics, and strong governance.
What This Means in Practice, and How Detection Helps
Translated into action, meeting Singapore's expectations means being able to answer, at every trusted touchpoint, whether the face or voice is real. That maps directly onto deepfake detection. At onboarding, it means screening selfies and documents for synthetic media and tampering, satisfying the MAS expectation to verify content authenticity. Alongside liveness, it means adding detection that catches the face swaps a liveness check passes, and pairing that with injection detection for the endpoint-level protection MAS calls out. For the vulnerability-testing expectation, a detection capability gives institutions something concrete to simulate deepfake attacks against. And extending detection to high-value transactions and voice channels addresses the additional-verification expectation and the call-center exposure.
This is where DuckDuckGoose, based in Delft, fits within a MAS-aligned stack. DeepDetector analyzes images and video for the signatures of synthetic media, catching AI-generated faces, injected feeds, and replay attacks beyond standard liveness, and it is built to integrate alongside eKYC workflows without adding friction for genuine customers. Its explainable output supports the governance and auditability that supervisory expectations increasingly demand, and it carries ISO 27001, SOC 2, and GDPR compliance. It complements, rather than replaces, the liveness, fraud-surveillance, and process controls the framework requires. For why liveness alone leaves a gap, see our guide to why liveness checks fail against deepfakes, and for the injection vector MAS names specifically, how injection attacks feed deepfakes into verification.
Table 3: How deepfake detection maps to what MAS expects.
Frequently Asked Questions
What does MAS require regarding deepfakes?
MAS's September 2025 information paper on deepfake cyber risks sets out expected measures for financial institutions: verify the authenticity of documents and digital content at onboarding, use liveness detection that covers deepfakes, add endpoint-level protection against injection attacks, run vulnerability tests that simulate deepfake attacks, integrate real-time deepfake detection into channels, and add verification for high-risk transactions. Binding obligations under MAS Notice 626 and the Shared Responsibility Framework sit alongside these expectations.
What is the Shared Responsibility Framework?
It is a MAS and IMDA framework, effective December 2024, that allocates responsibility for losses from a defined scope of phishing scams among financial institutions, telcos, and consumers, and requires payouts to victims where an institution breached its anti-scam duties. Those duties include real-time fraud surveillance and holding rapid account-draining transactions for at least 24 hours.
Does the Shared Responsibility Framework cover all deepfake scams?
No. It covers phishing scams with a digital and territorial nexus and excludes malware-enabled scams and authorized-payment scams such as investment or romance scams, where the victim intended the transfer. Since much deepfake fraud arrives through impersonation and authorized-payment routes, the SRF is one lever among several, which is why MAS pairs it with broader supervisory expectations.
Why is Singapore a target for deepfake fraud?
Because it concentrates wealth, has a dense financial sector, near-universal digital banking, and a regional-hub role, all of which raise the payoff per attack. Fraudsters have used deepfakes of executives and even national leaders in scams, and scam losses exceeded 913 million dollars in 2025, making it a priority for both attackers and regulators.
Is a deepfake-opened account just a fraud problem?
No, it is also a compliance problem. A synthetic identity or deepfake that defeats onboarding creates a fraudulent account that can constitute a KYC, anti-money-laundering, and counter-terrorism-financing failure under MAS Notice 626, exposing the institution to remediation costs and potential enforcement on top of the direct fraud loss.
Do MAS expectations cover voice and call centers?
Effectively yes, because the expectations extend to communication channels and high-risk transactions, and Singapore's major banks run large voice authentication operations. Voice cloning can replicate a customer or executive from seconds of audio, so the call center is part of the exposure, and detection and step-up verification apply there as well as at onboarding.
How can institutions meet these expectations?
By layering deepfake-aware detection across the customer journey: screening documents and selfies at onboarding, adding deepfake and injection detection alongside liveness, extending checks to high-value transactions and voice channels, testing systems against simulated deepfake attacks, and keeping explainable records for governance. This combines with the real-time fraud surveillance and transaction holds the framework already requires.
About the author
