> ## 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.
# Deepfake Detection
> How FaceSign detects deepfakes with 99% or higher accuracy using interactive liveness analysis.
FaceSign detects deepfakes, face swaps, and presentation attacks at a rate of 99% or higher, matching the **iBeta PAD Level 2** standard for presentation attack detection. This page explains how the system achieves that rate and why interactive liveness analysis defeats attacks that fool static checks.
## Why static liveness checks fail
Traditional liveness detection asks the user to take a selfie or hold still for a moment. These approaches are vulnerable to:
* **Pre-recorded video** replayed on a screen held in front of the camera
* **Real-time deepfake generators** that render a synthetic face over a live camera feed
* **3D-printed masks** and silicone replicas
* **Photo injection attacks** that bypass the camera entirely and inject a static image into the video stream
A single-frame analysis cannot distinguish a high-quality deepfake from a real face. The temporal dimension — how a face moves and responds over time — is what separates real from synthetic.
## Interactive liveness analysis
FaceSign uses **interactive challenges** during the `LIVENESS_DETECTION` node. Instead of asking for a static selfie, the avatar directs the user through randomized, unpredictable prompts that require real-time physical responses.
### What the system checks
| Analysis layer | What it measures | What it catches |
| --------------------------------------- | ------------------------------------------------------------------------------- | ----------------------------------------------------------------- |
| **Frame-by-frame temporal consistency** | Micro-movements, skin texture flow, lighting response across consecutive frames | Pre-recorded video, looped playback, frame injection |
| **Micro-expression timing** | Involuntary facial muscle activation timing after stimulus | Deepfake generators that render expressions with artificial delay |
| **Physiological response** | Pupil dilation, blink patterns, blood flow indicators visible through skin | Masks, silicone replicas, static images |
| **Interactive challenge response** | Physical actions in response to randomized, unpredictable prompts | Pre-rendered deepfakes that cannot respond to unknown challenges |
| **Depth and parallax** | 3D facial geometry from motion parallax | Flat screens, printed photos, 2D image injection |
### Why interactive challenges defeat deepfakes
Pre-rendered deepfakes are generated in advance. The attacker produces a synthetic video and replays it during the verification session. This approach fails against interactive challenges because:
1. **The challenge is randomized** — The avatar selects from a pool of prompts at runtime. The attacker cannot predict what will be asked.
2. **The response must be immediate** — The system measures the latency between the challenge and the user's physical response. Real-time deepfake generators introduce measurable processing delay.
3. **The response must be physiologically consistent** — When a real person turns their head, hundreds of micro-movements occur in facial muscles, skin folds, and lighting angles. Deepfake generators approximate these but introduce statistical artifacts detectable across multiple frames.
FaceSign's liveness detection is not a single check. It runs continuously throughout the `LIVENESS_DETECTION` node, analyzing every frame. The longer the interaction, the more data points the system collects and the higher the confidence.
## iBeta PAD Level 2
FaceSign's deepfake detection meets the **iBeta Presentation Attack Detection Level 2** standard:
| Metric | Requirement | FaceSign |
| ------------------------- | -------------------------------------------------------- | ---------------------------------------------- |
| **Attack detection rate** | Varies by attack type | 99% or higher across all tested attack vectors |
| **Tested attack types** | Printed photos, screen replay, 3D masks, video injection | All passed |
| **False rejection rate** | Must remain usable for legitimate users | Below industry threshold |
PAD Level 2 is the higher of the two iBeta testing levels. It includes advanced attack types (3D masks, video injection) that Level 1 does not test.
## Attack types and detection
| Attack | How it works | How FaceSign detects it |
| ---------------------- | --------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
| **Screen replay** | Attacker plays a video of the victim on a phone/monitor | Temporal analysis detects screen refresh artifacts; interactive challenges go unanswered |
| **Real-time deepfake** | Software generates a synthetic face over a live feed | Processing latency on challenge-response; micro-expression timing anomalies |
| **3D mask** | Physical mask worn by the attacker | Physiological signals absent (no blood flow, no pupil response); texture analysis |
| **Photo injection** | Image injected into the video stream bypassing the camera | No temporal variation; fails all interactive challenges |
| **Face swap app** | Mobile app that swaps faces in real-time | Edge artifacts at face boundary; inconsistent lighting between swapped face and body |
## Integration
Deepfake detection runs automatically in two node types:
* **`LIVENESS_DETECTION`** — Dedicated interactive liveness check with randomized challenges
* **`FACE_SCAN`** — Liveness analysis runs alongside biometric capture
You do not need to configure deepfake detection separately. Adding either node to your verification flow activates the full analysis pipeline.
## Next steps
FaceSign's crown jewel: detecting when a user acts under duress.
The full data flow from capture through HSM tokenization to results.
Add LIVENESS\_DETECTION and FACE\_SCAN nodes to your verification flow.