Catching Deepfakes Before They Cost You

Fraud teams at financial institutions noticed the shift before the researchers published about it. The attack patterns started changing around 2022. Liveness checks that relied on simple behavioral prompts began getting defeated by video injection attacks. KYC selfies started showing faces that passed a quick human review but failed closer inspection. Document verification pipelines started catching bank statements and salary certificates with professionally rendered layouts that no known institution had issued.

The cost of producing convincing synthetic fraud — synthetic faces, synthetic documents, video spoofs — collapsed. The cost of detecting it did not keep pace, at least not in an accessible form that fit inside a production onboarding pipeline.

The consequences are not abstract. Fraud losses in financial services globally run into the tens of billions of dollars annually. For a mid-scale lending platform, even a fraction of a percent of applications containing synthetic documents can represent significant exposure. More importantly, the regulatory dimension is real: compliance teams are increasingly expected to demonstrate technical controls against synthetic fraud, not just human review processes. "We had a human look at it" is not sufficient when the documents were fabricated at pixel level.

This is what 9thSense Shield Sense is designed to address.

Three Fraud Vectors, Three Detectors

Sophisticated fraud attempts typically combine multiple synthetic elements. A fraudster constructing a false identity will often submit a deepfake selfie, a synthetically generated supporting document, and attempt to defeat the liveness check with a video replay or camera injection — three separate vectors, all targeting the same onboarding flow simultaneously.

Shield Sense addresses each vector independently.

Deepfake detection identifies whether a submitted face image is genuine or synthetic. This covers AI-generated faces — images produced entirely by generative models, never photographed — as well as face-swap manipulations where a real person's face has been digitally replaced with another. The detector produces a verdict and a confidence level. When it identifies manipulation, it can also indicate which regions of the image show signs of synthetic content — useful both for automated decisions and for surfacing evidence to human reviewers.

Synthetic document detection analyzes whether a submitted document is a genuine physical document or a digitally fabricated one. The challenge here is that AI-generated documents have become convincing at the layout level — the fonts are correct, the letterhead matches, the template is accurate. Detection works on properties of the image itself: statistical characteristics of how the image was produced that differ between a real document scanned from physical paper and one rendered by a generative system. The detector produces a risk classification alongside its verdict, giving compliance teams a consistent signal to act on.

Face liveness verification confirms that a submitted face image is of a live person physically present at the camera, rather than a spoof. Spoof types the system detects include printed photo attacks, digital display replays, and camera injection (where a pre-recorded video is fed directly into the camera API to bypass the capture step entirely). Knowing which spoof type was attempted is operationally useful — it tells your fraud team which attack vector is being used against your onboarding flow, allowing you to respond.

These three checks are designed to work together. Catching a deepfake selfie without checking document authenticity leaves a gap. Running liveness without checking whether the underlying face is synthetic also leaves a gap. Shield Sense closes all three.

Fraud Detection as a Workflow Requirement, Not an Add-On

The most important architectural decision in Shield Sense is that fraud detection is built into verification workflows, not bolted on afterward.

In a typical point-solution approach, fraud detection is a separate service you call at some stage in your pipeline. It operates independently of your document verification. The results have to be correlated manually or through custom integration. The agent executing the workflow does not inherently know that a fraud check failed — your code has to wire that together.

In 9thSense, fraud checks are part of the agent's verification configuration. When you configure a KYC agent, you specify which fraud checks are required, when they run, and what the consequences of failure are. The agent treats a failed liveness check the same way it treats a missing document — it knows about it, it acts on it, it communicates about it to the applicant in plain language without exposing system internals.

This matters for the applicant experience as well as for fraud prevention. A liveness check failure that produces a confusing technical message erodes trust and increases support volume. An agent that says "The selfie you submitted could not be verified as a live photograph — please retake it in good lighting with your face fully visible" handles the failure gracefully and gives the legitimate applicant a path forward.

Agents can also be configured with different failure modes for different checks. A hard stop terminates the workflow immediately with no retry — appropriate for liveness and deepfake detection, where there is no legitimate reason for a genuine applicant to fail multiple times. A managed failure with limited retries is appropriate for synthetic document detection, where low-quality scans of legitimate documents can sometimes produce uncertain results. You configure the appropriate behavior for your use case.

PEP and Sanctions Screening: Real People, Not Just Fake Ones

Fraud detection is not only about synthetic media. A portion of the fraud risk in financial services KYC comes from real people who should not be onboarded: politically exposed persons requiring enhanced due diligence, individuals on sanctions lists, known fraudsters appearing under new identities.

9thSense Identity Sense handles this through face search across watchlist databases. When a face is submitted during onboarding, it is screened against databases that may include PEP lists, sanctions lists, internal fraud watchlists, and organizational databases. A match surfaces to your compliance team with the source database identified — so the analyst knows whether they are looking at a sanctions hit, an internal flagged account, or a PEP requiring enhanced due diligence.

This runs alongside document verification in the same case. Your compliance analyst sees the full picture: document extractions, fraud check verdicts, and identity screening results, all attached to the same case, not scattered across separate systems.

The Self-Hosted Advantage

Every classifier in Shield Sense runs on models within your infrastructure — either hosted by 9thSense in your designated region, or on your own servers if you self-host the intelligence layer.

This is not optional for many of our enterprise customers. KYC selfies and identity documents are among the most sensitive data a financial institution handles. Routing them to a third-party AI API creates data residency risks, compliance exposure, and dependency on a vendor's availability and pricing.

Self-hosted classifiers eliminate these concerns. Your customer's biometric data does not leave your environment. Your fraud detection availability is bounded by your own infrastructure, not a vendor's API status page. And you retain the ability to bring custom-trained models into the framework — if you have a fraud detection model tuned on your specific customer population, you can register it and route your tenant's checks through it instead of the built-in models.

The platform supports this by design. The intelligence layer is an internal API, and the model registry is how you configure which model handles which check for which tenant.

Compliance: What Auditors Expect

A documented, auditable fraud detection pipeline is increasingly what regulators expect from financial institutions — not just for external audits but for internal risk management.

Every fraud check in 9thSense produces a result that is attached to the case record. The check type, the verdict, the confidence level, the timestamp, and the agent action taken are all logged. When a compliance audit asks how you screen for synthetic documents in your onboarding flow, the answer is a case record, not a verbal description of a process.

For institutions operating under frameworks that require demonstrable technical controls against fraud — and this increasingly includes RBI-regulated entities, NBFCs, and any organization handling Aadhaar-based KYC — an auditable pipeline is not a nice-to-have.

Where Shield Sense Fits

Shield Sense is available on all 9thSense plans. The three fraud detection capabilities — deepfake detection, synthetic document detection, and face liveness verification — are part of the built-in capability set. They can be included in any agent's verification configuration without additional setup.

For use cases where you want to run fraud checks outside of an agent workflow — a standalone screening call, a batch verification job, an API integration with an existing pipeline — the platform's tool API supports direct invocation as well.

The goal is to catch fraud at the point of submission, before a fraudulent record is created, before a loan is disbursed, before an account is opened. That is the only point in the workflow where the cost of catching it is genuinely low.

After that, the costs compound.

Shield Sense docs → | Talk to us about fraud →