A production identity-verification service that decides whether a face-verification submission comes from a real, live person — or from a spoof or deepfake.
What it does#
- A multi-stage decision pipeline: quality gate → liveness gate → model
consensus, returning an
accept / review / rejectdecision with reasons. - Combines physical liveness cues (multi-frame RGB light-reaction checks) with an ensemble of deep models for deepfake detection.
- Served as an HTTP API for synchronous verification.
Tech#
- Models — convolutional networks (EfficientNet / Xception / CLIP-based) plus a gradient-boosting model, combined by consensus.
- Stack — Python, PyTorch, ONNX, FastAPI; end-to-end train → evaluate → deploy pipeline with automated retraining and reporting.
- Results — multi-model AUC ≥ 0.99 on held-out data; tuned to run within KYC latency budgets on CPU.