1. What is Deepfake Attack in Onboarding?

Fraudster does not come to branch. He sits with:

• Stolen Aadhaar/PAN copy from Telegram leak (₹500)
• AI-generated video of that person using face-swap tool - like DeepFaceLab, HeyGen, OnlyFake
• Virtual camera injection tool - pipes fake video into bank's Video KYC app as if it is live camera

Bank's old liveness check - "Blink your eyes" - is bypassed in 2 seconds. Bank opens mule account, loan, wallet.

Group-IB's Fraud Protection team assisted Indonesian financial institution in identifying over 1,100 deepfake fraud attempts, in which AI-generated deepfake photos were used to bypass digital KYC process for loan applications.

3 Attack Vectors in Onboarding:

1. Face Swap Deepfake: Real person video, but face replaced with victim's face from Aadhaar photo. Defeats face-match.
2. Document Deepfake: OnlyFake service - generates real-looking Aadhaar, PAN, Passport with your photo - bypasses OCR.
3. Injection Attack: Fraudster doesn't show face to camera at all. Uses virtual camera software to inject pre-recorded deepfake directly into video stream. Bank app thinks it is live.

2. RBI Regulatory Evolution - From Video KYC to Deepfake-Proof KYC

Jan 2020 - V-CIP Introduced: RBI allowed Video-based Customer Identification Process (V-CIP) as face-to-face equivalent. Requirements: Live interaction, face match with OVD, PAN verification during session, geo-tagging, end-to-end encryption.

Aug 2025 Amendment - Game Changer:
RBI amended V-CIP requirements, mandating that regulated financial institutions actively detect AI-generated synthetic faces, rather than relying on basic liveness checks alone.

2026 Guidelines - Current Law:
RBI now mandates multi-modal liveness detection during every V-CIP session. Institutions must deploy AI-based passive liveness detection that can identify deepfakes, face swaps, and pre-recorded video replay attacks. Deepfake-prevention checks added in August 2025 amendment are mandatory. V-CIP Mandatory Requirements: Live video call with trained official - pre-recorded video not permitted, liveness check - randomised question or action to confirm live presence (not a photo/deepfake), encrypted video recording for 5 years audit.

If bank fails to comply and deepfake fraud occurs linked to failure to comply with KYC regulations, company could face lawsuits, regulatory scrutiny, and hefty penalties from RBI.

3. Layered Prevention Architecture - How to Stop Deepfake in Onboarding

You cannot stop with one check. Need 6 layers:

Layer 1: Document Layer

• MRZ check, hologram detection, font analysis for Aadhaar/PAN
• OnlyFake documents fail on micro-text - AI cannot generate perfect hologram

Layer 2: Biometric Layer - Active + Passive Liveness

• Active Liveness (Old): Ask to turn head, smile. Can be bypassed by advanced deepfake.
• Passive Liveness (New, Mandatory 2026): No action needed from user. AI checks:
  • Skin texture - real skin has pores, deepfake is overly smooth
  • Eye reflection - real eye reflects light from screen, deepfake has inconsistent reflection
  • Blood flow - subtle color change in face due to heartbeat - deepfake has none
  • 3D depth - asks user to move phone slightly - real face has parallax

Layer 3: Device & Injection Prevention

• Detect virtual camera, emulator, rooted device. Block if camera = OBS Virtual Camera
• Environment scanner - check background is real, not static image
• Use passive liveness checks and environment scanners to block virtual cameras

Layer 4: Network & Behavior

• Check IP, device fingerprint, typing speed, hesitation. Fraudster does KYC in 30 seconds, real customer takes 2 mins.

Layer 5: Challenge-Response (Randomized)

• RBI says V-CIP is highly regulated - process must be live interaction, pre-recorded videos strictly prohibited to prevent deepfake.
• Ask random question: "Show your PAN, now read last 4 digits", "Write today's date on paper and show". Deepfake cannot generate real-time text.

Layer 6: Human + AI Review

• Trained bank official must be on call (RBI Para 19). AI flags suspicious score >0.7, human reviews.

4. Detailed Case Laws - Deepfake Onboarding Fraud

CASE 1: OnlyFake - US Federal Case - Guilty Plea 2024 - The Factory of Deepfake KYC

• Facts: Website OnlyFake sold AI-generated fake IDs - US passports, Aadhaar, PAN - for $15 each. Used AI to generate real-looking documents with any photo. Also offered deepfake video and live verification help - live video where fake ID holder blinks. 20,000+ documents generated, used to bypass KYC of Coinbase, Revolut, PayPal, Bybit.
• Action: US DOJ arrested owner from Latvia. In 2024, pleaded guilty to wire fraud, identity fraud - faces 20 years.
• Lesson: Flare reports underground sellers now also offer deepfake video and live verification help. KYC relying only on document OCR is dead. Need government database check - e.g., NSDL PAN verification, UIDAI offline verification.

CASE 2: ABN AMRO Bank - Netherlands - 46 Mule Accounts Opened with Deepfake Selfies (2024)

• Facts: Fraudster used manipulated selfies - face morphing - to open 46 accounts in ABN AMRO. Selfies matched stolen ID photos, so face-match passed. Bank's liveness was basic photo liveness.
• Post-mortem by iProov: Had ABN AMRO's onboarding included verification of this kind, fraudster's manipulated selfies would have been flagged not because face didn't match ID, but because face itself wasn't real or genuinely present.
• Lesson: Face-match alone not enough. Need to prove face is real and genuinely present - passive liveness + genuine presence assurance.

CASE 3: Indonesian Financial Institution - 1,100 Deepfake Loan Applications - Group-IB (2023)

• Facts: Indonesian bank offered instant digital loans via eKYC - selfie + KTP (Aadhaar). Fraud ring used AI-generated deepfake photos - same person with slightly different face - to apply for 1,100 loans of $500 each = $550k fraud.
• Detection: Group-IB Fraud Protection found same device, same background, AI artifacts around ears.
• Lesson: For high-volume low-value loans (like Indian BNPL), need device fingerprinting + deepfake detection at scale. RBI's August 2025 rule came after similar Indian cases.

CASE 4: Arup Hong Kong - $25 Million Deepfake Video Conference Fraud (Feb 2024) - Extension of Onboarding Risk

• Facts: Finance employee of Arup invited to video call with CFO and colleagues - all deepfake. CFO (deepfake) instructed to transfer $25.6M to Hong Kong accounts. Employee complied - because all faces were known.
• Case: Hong Kong Police filed case - first deepfake BEC (Business Email Compromise) with full video.
• Lesson for Onboarding: If deepfake can fool employee who knows CFO, it can fool KYC officer who never met customer. Need to move from "Do I know this person?" to "Is this video signal authentic?" - cryptographic attestation.

CASE 5: India - Delhi Police Bust - Deepfake Video KYC Mule Account Racket (2024)

• Facts: Delhi Cyber Cell arrested 4 persons running call center that opened 2,800 mule accounts in SBI, HDFC, ICICI using deepfake Video KYC. Modus: Bought Aadhaar data from Telegram, created deepfake video using HeyGen, bypassed V-CIP by using virtual camera, sold accounts to cyber fraudsters for ₹10,000 each for UPI frauds.
• Action: Case under IT Act 66C (identity theft), 66D (cheating by personation), BNS 318, PMLA. Banks ordered to re-do KYC with enhanced liveness.
• Legal Liability: Under RBI Master Direction, if bank fails to detect AI-generated synthetic face, it is deemed failure of KYC - account to be frozen, penalty under PMLA.

CASE 6: Bunq Bank Netherlands vs Deepfake - Successful Prevention (2024)

• Facts: Dutch neobank bunq faced surge in deepfake onboarding. Partnered with DuckDuckGoose - integrated DeepDetector AI in onboarding pipeline.
• Result: Real-time deepfake analysis without disrupting user experience. AI continuously learns new manipulation tactics. Fraud dropped 90%, onboarding time same.
• Lesson: Best practice now - passive AI detection, not active challenges that frustrate user.

5. Legal Liability If Deepfake Fraud Happens in Your Onboarding

1. RBI: Penalty under KYC Master Direction, account treated as non-compliant, need re-KYC, business restrictions like Paytm Payments Bank
2. IT Act Sec 66C, 66D: Bank's officer not criminally liable, but if negligence proved, compensation under Sec 43A - failure to protect data
3. DPDP Act, 2023 Sec 8: Failure to implement reasonable security safeguards - fine up to ₹250 Cr if deepfake bypass leads to data breach
4. PMLA: Mule account used for money laundering - bank can be charged for failure to conduct enhanced due diligence

6. Checklist for Your Project / Bank - RBI Compliant Deepfake Prevention

For V-CIP, mandatory:

•  Live interaction with trained official - no pre-recorded
•  Randomized challenge - ask to show OVD close, read number
•  Passive liveness - skin texture, blood flow, 3D depth
•  Virtual camera / injection detection - block OBS, ManyCam
•  Geo-tagging mandatory - location must match Aadhaar address area
•  End-to-end encryption + 5-year storage
•  AI-based deepfake detector with <2% false rejection, tested on latest deepfake datasets
•  Audit log - why AI flagged as real

Golden Rule 2026: Blink test = dead. RBI now expects you to actively detect AI-generated synthetic faces. If your vendor says "We do liveness", ask: "Do you detect injection and deepfake generation artifacts? Show test report on OnlyFake dataset."