What AI obligations does the RBI KYC Master Direction impose on V-CIP deployments?

The RBI Master Direction on KYC (DBR.AML.BC.No.81/14.01.001/2015-16, as amended through April 2023) explicitly mandates AI use in Video Customer Identification Process (V-CIP): 'Appropriate Artificial Intelligence (AI) technology can be used to ensure that the V-CIP is robust and it shall be ensured that face liveness/spoof detection is done with a high degree of accuracy.' This creates four specific AI governance obligations: (1) Face liveness detection with 'high degree of accuracy' — a model performance obligation that requires documented TPR/FPR thresholds in the model card, validated on a representative sample of Indian face types and lighting conditions; (2) Anti-spoof detection — the AI must detect both 2D photo/video replay attacks and AI-generated deepfake video attacks; (3) Face matching against Aadhaar OTP-linked biometric — the matching model's accuracy obligation is tied to the Aadhaar biometric stored at UIDAI, and mismatches must be handled by the regulated entity's fallback process; (4) Audit trail — V-CIP AI decisions (liveness pass/fail, face match confidence score) must be preserved for 8 years under the MD's record retention obligation. Under MRMF, the V-CIP AI model is a model used in a material decision (account opening / KYC compliance) and must be in the RE's model inventory with validation evidence.

What deepfake attacks have been documented against India's V-CIP and e-KYC systems?

Multiple incidents of AI-generated deepfake video attacks against India's digital KYC infrastructure have been documented. Attackers use AI face-swapping tools to overlay a synthetic face matching the PAN/Aadhaar document over a live video feed, bypassing face-matching components that compare the V-CIP video to the ID document photograph rather than a genuine liveness signal. The attack surface has three vectors specific to India's architecture: (1) DigiLocker-linked face matching — when the V-CIP pulls the reference photo from DigiLocker, the attack must generate a face that matches the Aadhaar photograph specifically; (2) e-NACH and payment authentication — banks using video KYC for e-mandate setup face the same deepfake attack surface outside account opening; (3) CERSAI and credit information company reporting — synthetic KYC identities created through V-CIP bypass can generate fraudulent credit records before detection. Under the MD, the RE is liable for KYC failures regardless of whether the failure originated in a vendor-supplied AI model, because V-CIP governance obligations fall on the regulated entity. The CERT-In April 2026 advisory specifically named vKYC deepfake bypass as one of the realized AI threat categories in India BFSI.

How does MRMF apply to V-CIP AI models and what documentation is required?

RBI MRMF June 2026 applies to V-CIP AI as a model used in a material compliance decision (KYC verification, account opening). The V-CIP AI model's MRMF documentation must include: (1) Model purpose and scope — V-CIP liveness detection and face matching; model coverage (which document types, face presentation conditions); (2) Performance metrics on validation dataset — TPR (true positive rate for genuine customers passing liveness), FPR (false positive rate for presentation attacks passing liveness), face match confidence distribution; the 'high degree of accuracy' standard in the MD requires documented thresholds — recommended TPR above 98%, FPR below 0.5% for a model used in account opening; (3) Adversarial testing — the model must be tested against AI-generated deepfake attacks, not only 2D replay attacks; the CERT-In April 2026 threat taxonomy provides the test categories; (4) Demographic performance analysis — MRMF's INDIC-BIAS framework requires testing for differential false rejection rates across demographic groups; a liveness model that fails more frequently for darker skin tones or older customers creates both compliance and fairness risk; (5) Third-party model governance — if the V-CIP AI is from a vendor, MRMF Section 7 applies; the RE cannot accept the vendor's own testing as satisfying its MRMF independent validation obligation; (6) PSI/CSI monitoring for ongoing performance — quarterly at minimum; presentation attack patterns evolve as attackers acquire better deepfake tools, requiring model update triggers.

What are the CKYC upload AI governance obligations under the RBI KYC Master Direction?

The Central KYC Records Registry (CKYCR), governed by SEBI and operationalized under the Prevention of Money Laundering (Maintenance of Records) Rules 2005, requires regulated entities to upload KYC records in the CKYCR format. AI models used to extract, classify, and validate KYC data before CKYCR upload have specific governance obligations. The RBI KYC MD requires that KYC records be complete, accurate, and updated — AI extraction errors that result in incomplete or incorrect CKYCR uploads create PMLA exposure for the RE. Three specific CKYC AI governance requirements: (1) Document classification accuracy — the AI model that classifies the KYC document type (PAN, Aadhaar, passport, voter ID) must maintain documented accuracy metrics; misclassification produces incorrect CKYCR category entries; (2) Data extraction accuracy — OCR and AI extraction of name, date of birth, address, and identification number must meet a documented accuracy threshold; extracted data that fails CKYCR validation rules creates upload failures and compliance gaps; (3) CKYC deduplication AI — when a customer presents as new but already has a CKYC record, the AI matching logic must correctly identify the existing record; failure creates duplicate CKYC entries with AML implications. Under MRMF, CKYC AI models are models used in compliance processes and must be in the RE's model inventory.

AI Governance · RBI KYC · V-CIP · Deepfake · MRMF · India BFSI

RBI KYC Master Direction V-CIP AI Governance: Deepfake Detection Obligations, Aadhaar e-KYC Model Performance Standards, and Third-Party Vendor Liability

The RBI KYC Master Direction does not merely permit AI in Video Customer Identification Process — it explicitly mandates it, with a requirement for "face liveness/spoof detection at a high degree of accuracy." This creates a model performance obligation that MRMF validates against. Deepfake attacks on India's V-CIP infrastructure are documented. No global AI governance competitor has mapped any of this.

Zytra Research Team · October 2026 · 13 min read

📖 Part of the India BFSI AI Governance Complete Guide 2026 — RBI, SEBI, IRDAI, NPCI & DPDP in one place.

India's Video Customer Identification Process (V-CIP) framework, introduced by the RBI KYC Master Direction amendment in January 2020, is now the primary channel for non-face-to-face account opening across India's banking system. Hundreds of millions of accounts have been opened through V-CIP. Every regulated entity — banks, NBFCs, payment system operators — must either operate V-CIP themselves or use a regulated service provider.

What makes V-CIP unique from an AI governance perspective is that the RBI KYC Master Direction (DBR.AML.BC.No.81/14.01.001/2015-16, as amended) does not merely permit AI use — it explicitly mandates it: "Appropriate Artificial Intelligence (AI) technology can be used to ensure that the V-CIP is robust and it shall be ensured that face liveness/spoof detection is done with a high degree of accuracy."

This language creates what MRMF calls a model performance obligation: the model must achieve "high degree of accuracy" in liveness and spoof detection, and that accuracy must be validated, documented, and monitored. The MD's record retention obligation (8 years for KYC records) extends to AI decisions made during V-CIP. The RE's liability for KYC failures under PMLA applies regardless of whether the failure originated in a vendor AI model.

This post maps every AI governance obligation in the V-CIP and KYC framework to MRMF and FREE-AI compliance requirements.

The V-CIP AI Obligation: What "High Degree of Accuracy" Requires

The phrase "high degree of accuracy" in the MD is not defined with specific thresholds. RBI has not issued a circular specifying, for example, that face liveness detection must achieve 98% TPR at 0.5% FPR. This absence is governance risk, not compliance comfort — it means the RE must set and document its own thresholds, and those thresholds will be evaluated by RBI examiners against the "high degree of accuracy" standard on a reasonableness basis.

For MRMF compliance, the V-CIP AI model's model card must document:

Performance metrics on the validation dataset: True Positive Rate (genuine customers correctly passing liveness), False Positive Rate (presentation attacks incorrectly passing as genuine), and Face Match Confidence threshold used for accept/reject
Test population composition: the validation dataset must cover Indian face presentation conditions — varied skin tones, lighting conditions, device cameras (front-facing smartphone cameras at varied quality levels), age groups, and face occlusion patterns common in India (dupatta, cap, spectacles)
Attack type coverage: 2D photo replay attacks, 2D video replay attacks, 3D mask attacks, and — critically for 2026 — AI-generated deepfake video attacks using face-swapping tools available at consumer prices
False rejection rate by demographic segment: INDIC-BIAS testing is required to ensure the liveness model does not disproportionately reject genuine customers from specific demographic groups; a liveness model with higher false rejection rates for older customers or specific skin tones creates both regulatory and fairness exposure

The recommended thresholds for a MRMF-compliant V-CIP AI model: TPR for genuine customers above 98.5%, FPR for presentation attacks below 0.3%, false rejection rate variance across demographic segments below 2 percentage points.

Deepfake Attack Surface for India's V-CIP Architecture

India's V-CIP architecture has three characteristics that shape the deepfake attack surface specifically:

Aadhaar OTP-linked biometric reference: V-CIP links the video session to an Aadhaar OTP — the face in the video must match the face in the Aadhaar record. An attacker needs to generate a deepfake that matches the specific Aadhaar photograph for the identity being impersonated. The UIDAI biometric database is not accessible for real-time face comparison by REs — the reference is the Aadhaar photograph on the card or DigiLocker. Deepfake generation models trained on a photograph of the target identity can now produce real-time video faces that pass face matching components comparing against a photograph.

DigiLocker document reference: When V-CIP pulls the KYC document from DigiLocker, the AI face-matching component compares the V-CIP video face against the photograph in the DigiLocker document. This creates a known-reference attack — the attacker has access to the reference photograph (from a leaked, purchased, or social-media-derived image of the target) and can train the deepfake model against that specific reference.

Regulatory liability regardless of origin: Under the MD, the V-CIP must be conducted by a "bank official" or, for outsourced V-CIP, an authorised vendor. In either case, the KYC obligations fall on the regulated entity. An RE that outsources V-CIP to a vendor and relies on the vendor's AI model without its own independent validation has not discharged its MRMF obligations — the MRMF Section 7 third-party model governance requirements apply.

MRMF Documentation Requirements for V-CIP AI

The V-CIP AI model sits in three MRMF categories simultaneously: a compliance model (used in a regulatory KYC obligation), a security model (detecting fraud/presentation attacks), and potentially a customer-affecting model (false rejection delays or prevents account opening). All three categories trigger MRMF obligations.

MRMF Section 3 (Model Validation) requires:

Independent validation — the same team that built the V-CIP AI cannot validate it; vendor-supplied V-CIP AI must be validated by an internal team or external validator independent of the vendor
Validation on a population that reflects the RE's actual V-CIP customer base — not the vendor's generic benchmark dataset
Adversarial testing with current attack typologies — a validation conducted in 2023 that did not include AI-generated deepfake attacks requires update; the attack landscape has materially changed

MRMF Section 6 (Ongoing Monitoring) requires:

Monthly tracking of V-CIP completion rate, failure rate, and rejection reason distribution — model degradation may first appear as rising rejection rates or increased false rejection rates in specific customer segments
Incident logging for V-CIP sessions that resulted in subsequently discovered fraudulent accounts — post-hoc identification of deepfake-bypassed sessions to update the model's adversarial test dataset
Quarterly PSI monitoring for liveness score distribution — a distribution shift may indicate that the model is being probed at scale by attackers learning to optimize their deepfake inputs toward the model's accept region

CKYC Upload AI Governance

Regulated entities are required under the Prevention of Money Laundering (Maintenance of Records) Rules 2005 and the RBI KYC MD to upload KYC records to the CKYCR within 10 days of establishing a customer relationship. For V-CIP-originated accounts, the CKYC upload relies on AI-extracted data from the KYC documents presented during V-CIP.

Three AI governance obligations for CKYC upload AI:

OCR accuracy for CKYC fields: The AI OCR model extracting name, date of birth, address, PAN, Aadhaar number from V-CIP documents must achieve documented accuracy above a threshold on the CKYC mandatory fields. Extraction errors result in CKYCR validation failures or incorrect records — both create PMLA compliance gaps that are attributable to the RE.

Document classification: The AI model classifying the document type (OVD category under the MD) must correctly identify the document to ensure the right CKYC template is applied. Misclassification causes field mapping errors that produce structurally incorrect CKYC records.

CKYC deduplication: The AI matching logic that checks for an existing CKYC record for the customer must operate at a recall level that prevents duplicate CKYC records — duplicate records create AML blind spots where the aggregate exposure across a customer's accounts cannot be viewed through a single CKYC identifier.

Third-Party V-CIP Vendor Governance Checklist

☐ Vendor V-CIP AI in RE's MRMF model inventory as third-party model; model card entry complete
☐ Independent validation conducted by RE or commissioned external validator — vendor's own performance statistics are not sufficient; vendor must provide model documentation sufficient for independent validation
☐ Validation dataset composition documented: Indian face presentation conditions, current attack types including AI-generated deepfakes, INDIC-BIAS demographic segments
☐ Performance thresholds set and documented in model card: TPR, FPR, face match confidence threshold; thresholds reviewed against MD "high degree of accuracy" standard with documented rationale
☐ Change notification: vendor contract requires notification before any model update; RE assesses update impact before continuing reliance
☐ PSI monitoring: V-CIP liveness score distribution tracked; quarterly at minimum; threshold for investigation and revalidation defined
☐ DPDP compliance for biometric data: V-CIP biometric processing is sensitive personal data under DPDP Act; consent obtained; data retention per MD 8-year requirement with DPDP erasure reconciliation plan
☐ Incident response: process for identifying and logging V-CIP sessions that resulted in fraudulent accounts; AI model update triggered by confirmed deepfake bypass incidents
☐ CKYC AI governance: OCR accuracy thresholds set; document classification accuracy tracked; CKYC deduplication recall rate monitored
☐ RBI examination readiness: V-CIP AI model documentation package (MRMF model card, validation report, monitoring dashboard) ready for examination on request