What Is Deepfake Detection for KYC in India, and Where Does It Fit Aadhaar eKYC and V-CIP?

What is deepfake detection for KYC in India?

Deepfake detection for KYC in India is an automated check on the face media a customer submits, asking whether that media shows signs of face swapping, reenactment, full synthesis or pixel-level editing. It runs on the same selfie, liveness clip or video frame that the face-match step already uses, and it returns a risk signal that the bank's policy turns into approve, step up or review.

Indian onboarding has several routes where a face is the deciding evidence. Aadhaar face authentication compares a live capture against the UIDAI record. Digital KYC under the RBI directions captures a live photograph of the customer at the point of onboarding. V-CIP is a live audio-visual interaction between a trained bank official and the customer, recorded and geotagged. Each route assumes the face on camera belongs to a real person who is present.

Generative tools break that assumption in two ways. An attacker can present a deepfake to the camera on a second screen, or inject synthetic frames so the app never sees a real sensor feed. Liveness and spoof detection was designed mainly for the first case; deepfake detection examines the content of the frames whichever way they arrived.

Volume makes this a practical issue. UIDAI reported 28.29 crore Aadhaar face authentication transactions in November 2025 against 12.04 crore in November 2024, and 47.19 crore e-KYC transactions in the same month. Face checks are now a high-volume production control, which is the condition under which automated attacks pay off.

Do Indian banks need liveness and deepfake detection APIs?

Indian banks running V-CIP need liveness or spoof detection by rule, and deepfake detection is the most direct way to meet the parts of the rule that liveness alone does not cover. Paragraph 27(1)(v) of the RBI Commercial Banks KYC Directions, 2025 says the V-CIP application must have components with face liveness or spoof detection as well as face matching technology with a high degree of accuracy. Paragraph 27(2)(i) also expects the official to be capable of carrying out a liveness check and detecting any other fraudulent manipulation.

The directions do not use the word deepfake. They do say the bank may use appropriate AI technology to strengthen V-CIP, and paragraph 27(1)(vii) asks banks to update technology, software and workflows based on detected, attempted and near-miss cases of forged identity, and to report any detected forged identity through V-CIP as a cyber event. A face swap that passes liveness is exactly the kind of near-miss that clause anticipates.

Liveness and deepfake detection matter for Indian banks because they answer different questions. Liveness asks whether a live person is in front of the camera. Deepfake detection asks whether the face in the frames is the real, unaltered face of that person. A real-time face swap streamed by a live attacker can satisfy the first question and fail the second.

How do deepfake risks affect Indian KYC and AML programs?

Deepfake risks reach AML programs through the accounts that slip past onboarding. An account opened with a synthetic or borrowed face can be used as a mule account to receive and move fraud proceeds, and the transaction monitoring team inherits a customer record that was never genuine. The sibling guide on mule account enrollment with synthetic media covers that pattern in detail.

Re-KYC and account recovery carry the same exposure. Periodic updates, mobile number changes and device rebinding often rely on a fresh selfie or video, and an attacker who has taken over an account needs to pass only that one check.

Impersonation also targets customers directly. On 19 November 2024 RBI warned the public that fake videos of its Governor promoting investment schemes were circulating on social media, and stated that it does not give such investment advice. Scams of that kind feed fraud into banks through authorised payments rather than onboarding, so they belong in the same risk register.

Platforms that host synthetic content have separate duties. India's amended IT intermediary rules, in force from 20 February 2026, address labelling and provenance for synthetically generated information; banks should treat them as context, since the V-CIP obligations come from RBI.

How does document deepfake detection fit Indian eKYC?

Document deepfake detection fits Indian eKYC mostly at the edges, because the core identity data is already verified against issuers. In V-CIP, Aadhaar is verified offline through an XML file or Aadhaar Secure QR Code that must not be older than three working days, and PAN details are verified against the issuing authority's database. A printed copy of an e-document, including e-PAN, is not valid for V-CIP.

Those controls check that the data is genuine. They do not prove that the face printed on an officially valid document image belongs to the person on camera, and uploaded document images in non-V-CIP journeys can still be edited or generated. The attack that matters is often photo substitution: a real document number with a synthetic or swapped face.

NeuroVerify does not analyse ID documents. It checks the live face side of the match, which is where photo substitution is meant to be caught. When an analyst wants to examine a suspect document image itself, it can be uploaded to NeuroDetect as part of a forensic case with region-level views and a downloadable report. The broader mechanics are covered on the document deepfake detection for KYC page.

How does Neural Defend strengthen an existing Indian KYC stack?

Neural Defend strengthens an Indian KYC stack by adding one media check after capture and before the decision, without replacing the liveness, face match, Aadhaar or PAN steps already in place. The bank's capture SDK or V-CIP platform sends the selfie, frame or clip to the NeuroVerify API and receives one unified risk score with signal-level evidence.

One request runs six signals: Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance. Eligible image requests return in under 200 ms p95 under defined test conditions, so the check can sit inside a synchronous onboarding step. Mobile SDKs cover Android, iOS and Flutter; backend SDKs cover Python, TypeScript, Go and Java; there is also a REST API.

Data residency is usually the first question from Indian banks. NeuroVerify can be deployed as a managed cloud API, in a private cloud or VPC, as a secure container, on premises or air-gapped, with the option scoped during evaluation. Neural Defend holds ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II certifications.

On detection quality, the models build on ARCAS 1B, Neural Defend's deepfake detection foundation model; the ARCAS paper reports pooled accuracy of 99.99% on GenImage and 97.65% on OpenFake. On the public VendorBench-100 benchmark of 100 adversarial still images, Neural Defend ranked first of 36 systems with an MCC of 0.876 against 0.646 for the next vendor. One VendorBench author also co-authored the ARCAS paper, the corpus and results are public, and the benchmark does not test video or injection.

Where detection helps, where it stops, and what to keep as evidence

Detection is strongest against generated and edited faces: swaps, reenactment, synthetic faces and retouched selfies. It is weak against a genuine, unedited recording of the real customer replayed into the session, which is why RBI's variable questioning and session controls still matter. It also cannot tell whether a real person is acting under coercion or renting out their identity.

Treat each score as input to a policy. Route high-risk results to a trained official or a fresh V-CIP session, and record why an exception was granted.

For audit, keep the request identifier, model version, signal-level results, the decision and the official's credentials next to the V-CIP recording. That record supports the cyber event reporting in paragraph 27(1)(vii) and lets the bank show how near-miss cases changed its thresholds.