How Can Indian Banks Use Deepfake Detection in V-CIP, Onboarding and Phone Banking?

How Can Indian Banks Use Deepfake Detection?

Indian banks use deepfake detection at the points where a remote face or voice decides whether an account opens, a device is trusted or an instruction is accepted. In practice that means five places: the V-CIP video call, selfie-based digital onboarding, periodic re-KYC, re-binding a mobile banking app to a new device, and phone banking or relationship manager calls.

V-CIP carries the most regulatory weight because RBI prescribes how it must work. Selfie onboarding and re-KYC carry the most volume. Device re-binding and phone banking are where account takeover attempts concentrate, since an attacker with a customer's credentials still has to pass a face or voice check to move money.

What Do RBI's V-CIP Rules Require, and Where Does Deepfake Detection Fit?

RBI's Know Your Customer Directions for commercial banks, dated 28 November 2025, set out the V-CIP technology and procedure in paragraph 27. The application must include face liveness or spoof detection and face matching technology with a high degree of accuracy, though specific gestures such as blinking are not mandatory. Recordings must carry live GPS coordinates and a date and time stamp, data must be encrypted end to end, and connections from IP addresses outside India or from spoofed IP addresses must be blocked.

The procedure rules matter as much as the technology. Only specially trained officials may conduct V-CIP, they must vary the questions and their sequence to establish that the interaction is real-time and not pre-recorded, and the bank must reject the account if it sees prompting at the customer's end. All data and recordings must be stored in systems located in India.

Deepfake detection fits inside the spoof detection requirement. Liveness answers whether a live person is in front of the camera; a deepfake detector answers whether the face on the screen shows signs of being generated, swapped or manipulated. Running both on the same frames gives the V-CIP official two independent signals before they make the call.

Why Is Injected Video the Main Threat to V-CIP?

RBI's requirement to prove real-time interaction targets the oldest attack: playing a recording. Generative tools now allow a newer one, where a fraudster runs a face-swap model on a laptop and feeds the output into the V-CIP session through a virtual camera driver. The face moves, answers questions and responds to prompts, so varied questions alone do not expose it.

Three checks work together against this. Device and capture checks look for virtual cameras, emulators and modified apps. Liveness checks look for signs of a physical presence. Deepfake detection looks for the traces a face-swap or generative model leaves in the pixels. Our guide to virtual camera injection attacks explains how the first of these works.

Network quality complicates all three. V-CIP calls in India often run over mobile data with variable bandwidth, and heavy compression removes some of the fine detail detectors read. Banks should check scores on still frames captured at the highest available quality and ask for a recapture when quality is too low, rather than treating a blurry frame as a pass.

How Does Voice Deepfake Detection Fit Indian Financial Services?

Voice deepfake detection fits the phone banking, contact centre and relationship manager channels where Indian banks still accept spoken instructions. A cloned voice can be built from a short public clip, and the FBI warned in December 2024 that criminals use cloned audio to impersonate account holders and gain access to bank accounts.

The most useful placement is on calls that change something sensitive: a new registered mobile number, a card or limit change, a payee addition, or a request to resend credentials. Score the caller's audio, and when the score is high, move the request to a verified channel such as the mobile app or a branch visit instead of trying to judge the voice by ear.

Impersonation also reaches customers directly. On 19 November 2024 RBI cautioned the public that deepfake videos of its Governor were circulating on social media claiming RBI support for investment schemes, and stated that RBI does not give such advice. Banks can repeat that message in their own channels and add warnings and cooling-off periods to first payments to new beneficiaries.

How Should Banks Record Deepfake Findings for Fraud Risk Management?

RBI's revised Master Directions on Fraud Risk Management, issued on 15 July 2024, ask regulated entities to run governance, prevention, early warning, investigation and reporting frameworks. They do not name deepfakes, but a deepfake finding is evidence that has to fit into that framework.

Store the detector score, the signals behind it, the model version and the reviewer's decision with the V-CIP recording or onboarding record, inside India as the KYC directions require for V-CIP data. That record supports early-warning reviews when an account later shows mule behaviour, and it gives investigators the reason a session was allowed or stopped.

How Neural Defend Supports Indian Banks

Neural Defend's NeuroVerify checks a face image, video or audio clip in one API call and returns one risk score with evidence from 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 at p95 under defined test conditions, so frames from a V-CIP call can be scored while the official is still speaking with the customer.

Banks can deploy NeuroVerify in a private cloud or VPC, as a secure container, on-premises or air-gapped, which lets media stay inside infrastructure the bank controls. Neural Defend holds ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II. NeuroVerify sits next to your V-CIP and identity verification vendors, and it does not analyse ID documents. For phone channels, NeuroCall provides call-level voice deepfake detection and is currently in Beta.

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. The benchmark does not test video, audio or injection attacks, and one author also co-wrote Neural Defend's ARCAS 1B paper, so banks should test on their own V-CIP recordings. Our comparison of the top deepfake detection APIs in India sets out the alternatives.