Tencent Cloud AI FaceID Fake Face Detection

Problem to Solve

In face verification, remote onboarding, and identity checking, a system must not only confirm that a face is present, but also decide whether the face may come from screen replay, AI face swapping, batch black-market material, or watermark tampering. Basic face detection often answers only “is there a face,” while real fraud workflows need signals for AIGC face swaps, high-resolution replay, liveness spoofing, and other attack patterns. TencentCloud FaceID DetectAIFakeFaces is aimed at these anti-attack use cases, extending detection to single face images and short video clips and returning risk levels plus attack types that can be used in business decisions.

How the Skill Works

The skill wraps the Tencent Cloud FaceID AI fake-face detection API and gives Python callers a straightforward input-output path:

• Input: FaceInput accepts a local image/video path or a Base64 string; FaceInputType is optional, and local files can be inferred from the file extension.
• Media constraints: images support jpg/png, with Base64 size ideally under 3MB; videos support mp4/avi/flv, ideally 2–5 seconds, around 480×640, with Base64 size under 8MB.
• Detection scope: it covers AIGC face-swap detection, replay detection, black-market attack detection, watermark detection, and related liveness spoofing risks in the result model.
• Output shape: it returns JSON with risk levels such as Low, Mid, and High, plus concrete attack types, making it easier to feed into risk rules, manual review, or blocking logic.

From an engineering perspective, its value is not to replace liveness capture, but to add a second-layer anti-forgery check on face media that has already been collected: for example, a replayed ID photo, a short video with synthesized face swapping, batch signups using similar black-market clips, or images with suspicious watermark traces.

Boundaries and Notes

This skill fits environments where the caller already has Tencent Cloud FaceID permissions and lawful consent for face data. It focuses on anti-forgery detection, not general image classification or face matching; if the product requires strict identity consistency, it should be combined with face comparison, liveness capture, and other risk signals. Image/video formats, duration, resolution, and Base64 size are constrained, so integrations should handle preprocessing and fallbacks. Face data is personal information, so local regulations and user consent must be respected.