SHIELD-SISA weights
These are the weights for SHIELD-SISA, a native C++/CUDA engine and live service that detects audio and video deepfakes together in real time. The engine, its API and the build instructions are at https://github.com/tlstngud/shield-sisa (Apache-2.0).
Non-commercial demo / research use only (CC BY-NC 4.0).
Files
| file | model | notes |
|---|---|---|
shield_v.weights |
SHIELD-V video model, 9.88 M parameters, FP32 | MobileNetV3-Large frame encoder → GDN-A → shared GDN-B ×2 → SISA → head; reads 20 face-crop frames of 320×180 (10 s at 2 fps) |
shield_v.weights.json |
sidecar | decision threshold (score ≥ 0.26436 → fake), SHA-256 of the weights, tensor table |
shield_v.weights.cpu.cert, shield_v.weights.cuda.cert |
certificates | allow the reduced-precision modes (CPU BF16; CUDA FP16/BF16) for exactly these weights and this threshold |
audio/shield.weights |
SHIELD-A audio model, FP32 | reads 10 s of 16 kHz speech; run it with the stream low-pass enabled (the service default) |
audio/shield.weights.json |
sidecar | SHA-256 |
The native engines check the SHA-256 values at load time. If a file has been modified, they refuse to load it or fall back to exact FP32.
Usage
git clone https://github.com/tlstngud/shield-sisa && cd shield-sisa
./build.sh
hf download koreashin/shield-sisa --local-dir build
build/service/demo_av --video clip.mp4 --audio clip.mp4 --backend cpu
Video model results (RTFS face-swap benchmark, val + test, 840 videos)
All numbers use the exact FP32 tier at the threshold above.
| split | AUROC | false alarms (Real) | detected (Fake) |
|---|---|---|---|
| val | 0.9765 | 21 / 211 | 206 / 216 |
| test | 0.9908 | 21 / 232 | 180 / 181 |
The certified reduced-precision modes stay within these limits of exact FP32, per split and per fake method:
- AUROC drops by at most 0.002;
- the Real false-positive rate rises by at most 1 pp;
- the Fake true-positive rate falls by at most 1 pp.
Limitations
- The video model was trained and evaluated on face-swap methods (inswapper, uniface) and one dataset. It may not generalise to other generators, heavy compression, or unusual lighting and poses.
- Strong re-encoding or a different colour-space decode can shift the scores of real videos.
- When no face is detected, the model scores the whole frame.
- Do not use these models as the sole basis for decisions about people.
License
CC BY-NC 4.0. The engine code is Apache-2.0.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support