> ## Documentation Index
> Fetch the complete documentation index at: https://docs.facesign.run/llms.txt
> Use this file to discover all available pages before exploring further.
# Coercion & Duress Detection
> FaceSign's crown jewel: detecting when a user is acting under duress during verification.
Coercion detection is what separates FaceSign from every other authentication product. Most systems answer one question: **is this the right person?** FaceSign answers two: **is this the right person, and are they acting freely?**
This is FaceSign's primary differentiator. No call center agent, OTP code, or selfie-based verification can detect when a user is being coached, threatened, or manipulated into authorizing a transaction.
## Why coercion detection matters
Authorized push payment (APP) fraud — where the legitimate account holder is tricked or forced into making a payment — is one of the fastest-growing fraud categories globally. In these attacks:
* The account holder passes every traditional verification check (they are who they say they are)
* OTP codes are entered correctly by the actual user
* Call center agents hear the real account holder's voice
* The bank authorizes the transaction because every signal says "legitimate"
The missing signal is **state of mind**. The user is present, verified, and authenticated — but they are acting under duress, coaching, or social engineering. Traditional authentication cannot detect this.
## Three signal categories
FaceSign's coercion detection engine analyzes three categories of signals throughout the verification session. These run across all nodes, not as a separate step.
### Vocal signals
| Signal | What it indicates |
| ----------------------------- | --------------------------------------------------------------------------------------------------------------------- |
| **Stress frequency patterns** | Elevated fundamental frequency and vocal tremor that correlate with psychological stress |
| **Speech cadence changes** | Unnatural pauses, rushed speech, or irregular rhythm suggesting external coaching |
| **Hesitation markers** | Excessive filler words, false starts, and self-corrections before answering |
| **Vocal affect mismatch** | Emotional tone inconsistent with the content of the response (e.g., stress markers while claiming everything is fine) |
### Visual signals
| Signal | What it indicates |
| ---------------------------- | --------------------------------------------------------------------------------------------------------------- |
| **Gaze aversion** | Repeated glances away from the screen, suggesting reading from a script or checking with someone off-camera |
| **Micro-expression leakage** | Involuntary facial expressions (fear, anxiety, disgust) that flash before a controlled response |
| **Second-person detection** | Reflection analysis and peripheral motion detection that identifies another person present in or near the frame |
| **Facial tension patterns** | Sustained muscle tension inconsistent with a relaxed, voluntary interaction |
### Behavioral signals
| Signal | What it indicates |
| ----------------------------- | ----------------------------------------------------------------------------------------------------------------- |
| **Response timing anomalies** | Delays before answering that suggest consulting someone else, or responses that are too immediate (pre-rehearsed) |
| **Answer consistency** | Contradictions between responses given at different points in the conversation |
| **Compliance patterns** | Unusual eagerness to skip security steps or rush through verification |
| **Cognitive load indicators** | Signs that the user is simultaneously processing external instructions while responding to the avatar |
Coercion detection runs continuously across the entire session. Conversational nodes (`CONVERSATION`) provide the richest signal because they require the user to speak freely, but vocal and visual analysis runs during every node type.
## Use cases
### Elder financial abuse prevention
An elderly account holder is coached by a caregiver or family member to authorize a wire transfer. The account holder passes liveness detection — they are real and present. But vocal stress analysis detects elevated anxiety, gaze tracking shows repeated glances to someone off-camera, and response timing reveals delays consistent with external coaching.
### Coached wire transfer fraud
A victim receives a call from a scammer impersonating their bank. The scammer instructs the victim to authorize a "security transfer" through the real bank's verification flow. The victim follows the instructions, but their speech cadence is unnaturally deliberate (reading scripted responses), and micro-expression analysis detects fear signals that contradict their verbal cooperation.
### Romance scam authorization
A victim is emotionally manipulated into transferring money to a fraudster. During verification, behavioral analysis detects cognitive load indicators — the victim is processing conflicting emotions while trying to present a calm, voluntary front. Hesitation markers and vocal affect mismatch flag the session for review.
### PSD3 Strong Customer Authentication
Under PSD3, payment service providers face liability for coached transfer fraud. FaceSign's coercion detection provides an auditable signal that the user was — or was not — acting under duress at the moment of authorization. This evidence supports regulatory compliance and liability defense.
## How results are delivered
Coercion detection results are included in the session report delivered to your webhook. The report contains:
| Field | Description |
| ----------------------- | ---------------------------------------------------------------------------- |
| **Coercion risk score** | Numeric confidence score (0–100) indicating the likelihood of duress |
| **Signal breakdown** | Which signal categories contributed to the score (vocal, visual, behavioral) |
| **Flagged moments** | Timestamps within the session where specific indicators were detected |
| **Recommended action** | Suggested next steps based on the risk level (approve, review, block) |
Your application decides what to do with the score. FaceSign provides the signal — you set the threshold and the response.
## Next steps
Interactive liveness analysis with 99% or higher accuracy.
Six parallel AI models and HSM tokenization.
PSD3 Strong Customer Authentication and regulatory alignment.