Deepfake Detection
Deepfake detection is an additional layer of defense against injection attacks using face images. The feature is provided by Digital Identity Service.
Functionality in Digital Identity Service
The server provides the API endpoint POST customers/{id}/liveness/evaluation/extended, which allows you to request deepfake detection on liveness images. This check can also be performed on a selfie image if that image is linked as a passive liveness photo..
Recommended usage
Deepfake detection serves as a second line of defense. The primary defenses should be:
- Liveness (Presentation Attack Detection)
- Video Injection Detection It is recommended to apply deepfake detection only to photos that have already been evaluated as bona fide by both Liveness and Video Injection Detection.
If Video Injection Prevention is not used in the process (e.g., the image is not obtained using Innovatrics’ auto-capture components), deepfake detection is highly recommended in combination with Liveness. However, deepfake detection alone is not a sufficient substitute for Video Injection Prevention, as non-deepfake images can still be injected.
Thresholds for Deepfake detection
The following table shows the measured thresholds for specific working points, based on Innovatrics’ dataset of genuine photos and deepfakes created by generative AI models not used during development. Metrics are explained in the Liveness documentation.
Digital Identity Service 1.55.0 and above
| Use case type | Threshold | Performance |
|---|---|---|
| Convenience (minimum rejected attempts) | 0.781 | 4.4% APCER @ 1% BPCER |
| Balanced (equal error rate) | 0.818 | 1.7% both APCER & BPCER |
| Security (minimum accepted frauds) | 0.834 | 2.3% BPCER @ 1% APCER |
Known and Unknown Generative AIs
The feature detects:
- images fully generated by AI
- face swap images
The detection mechanism is designed to work even against images generated by AI models that were unknown or did not exist at the time of release.