Deepfake Detection
Deepfake detection refers to the technologies and methods used to identify images, videos, or audio generated or manipulated by artificial intelligence. In the context of identity verification, it aims to counter fraud attempts using synthetic faces.
Deepfakes pose a growing threat to remote identity verification systems. Using Generative Adversarial Networks (GANs) and diffusion models, it is now possible to generate photorealistic face images, swap faces in real-time video, or clone a voice from just seconds of recording. These techniques can be used to defeat face matching and liveness detection systems.
Deepfake detection solutions analyse several types of artefacts: pixel-level inconsistencies (blurred edges, abnormal textures), physiological anomalies (absent blinking, unnatural facial asymmetries), temporal artefacts in videos (flickering, distortions between frames), and image metadata (traces of AI generation in EXIF data).
The race between deepfake creators and detectors is constant. The most advanced solutions combine multiple detection models trained on diverse datasets, multimodal analysis (image + audio + behaviour), and continuous updates to adapt to new generation techniques. iBeta Level 2 certification now includes specific tests against deepfake attacks.
Regulations
Real-world examples
- 1.A fraudster attempts to open an online bank account using a real-time deepfake to pass the face matching step. The system detects micro-artefacts around the facial contours and inconsistencies in eye blinking patterns.
- 2.An electronic signature platform identifies a deepfake video during an identity verification journey: frequency analysis reveals compression patterns incompatible with an authentic camera capture.
- 3.An insurer detects a life insurance fraud attempt using a deepfake video of the purported policyholder. Temporal analysis of the video reveals micro-distortions typical of generative models between successive frames.