moyuan10086/aigc-safety-models

AIGC Safety System Models

This repository contains the local model artifacts used by the AIGC Safety System. It combines the project-trained Deepfake detector with the local retrieval, face-detection, and Shadow-audit models.

Included models

Path	Model	Role	Origin / license note
deepfake/model-epoch6.ckpt	AIGC Safety DFDet Multi-Dataset PEFT	Project Deepfake detector	Project checkpoint trained on DF40 + DeepFakeFace; use under the project license and dataset terms
deepfake/face_detection_yunet_2023mar.onnx	YuNet	Face detection and five-point alignment	OpenCV Zoo model; retain upstream attribution and license
embedding/paraphrase-multilingual-MiniLM-L12-v2/	Multilingual MiniLM	RAG semantic retrieval	Sentence-Transformers model; see its upstream model card and license
shadow/hybrid_safety_model_xgboost_color.json	Hybrid Safety XGBoost	Out-of-band Shadow audit	Local project/vendor artifact; treat serialized model data as trusted code and verify its digest

The repository also documents models that are part of the production system but are not distributed here:

• qwen3guard-gen-0.6b: called through the configured OpenAI-compatible Qwen3Guard service.
• singguard-nsfa-0.8b: called through the configured OpenAI-compatible SingGuard service.
• MLLM and text-generation models: selected by MLLM_MODEL and CHAT_MODEL_NAME and supplied by an external gateway.
• PaddleOCR and optional NudeNet models: installed or downloaded by the target runtime according to their own distribution terms.

Their weights remain on the authorized inference or deployment host and are not copied into this public repository.

Deepfake detector

• Backbone: CLIP ViT-L/14 visual encoder
• Adaptation: parameter-efficient LayerNorm tuning with a LinearNorm binary head
• Training data: DF40 + DeepFakeFace
• Checkpoint: epoch 6 (model-epoch6.ckpt)
• Task: per-face real/fake Deepfake detection

The checkpoint is intended for use with the inference code in moyuan10086/aigc-safety-system. It is not a general-purpose authenticity guarantee and should be combined with review and provenance evidence.

Files

• deepfake/model-epoch6.ckpt: PyTorch Lightning checkpoint for the project inference service.
• deepfake/face_detection_yunet_2023mar.onnx: YuNet face detector used by the preprocessing pipeline.
• embedding/paraphrase-multilingual-MiniLM-L12-v2/: Sentence-Transformers snapshot used by the RAG service.
• shadow/hybrid_safety_model_xgboost_color.json: XGBoost Shadow model used only for observation and evaluation; it never overrides the production guardrail verdict.

Usage

Download the files locally and set the application paths, for example:

DEEPFAKE_MODEL_PATH=/models/aigc-safety-models/deepfake/model-epoch6.ckpt
DEEPFAKE_FACE_MODEL_PATH=/models/aigc-safety-models/deepfake/face_detection_yunet_2023mar.onnx
GUARDRAIL_XGBOOST_SHADOW_MODEL_PATH=/models/aigc-safety-models/shadow/hybrid_safety_model_xgboost_color.json

Configure the Sentence-Transformers cache to use embedding/paraphrase-multilingual-MiniLM-L12-v2/. Keep downloaded artifacts outside the application Git repository.

Integrity

The project checkpoint SHA-256 is:

model-epoch6.ckpt  fd0b967e9ab6a19aa0127c5c2301b3883ff8013b1a5d7af282e29737132c1fb9

Recompute hashes after download and compare them with the release metadata before enabling a model in production.

License and attribution

The project checkpoint is released under the repository license where permitted. Dataset terms and the CLIP base model terms remain applicable. Review the upstream licenses before redistribution or commercial use.