Official PyTorch implementation of SSAF (Skeleton-assistant Standardization and Authentication Framework), a novel framework for robust and efficient dynamic hand gesture authentication in real-world uncontrolled scenarios.
Yufeng Zhang, Xilai Wang, Wenwei Song, and Wenxiong Kang
IEEE Transactions on Information Forensics and Security (TIFS), 2026
Dynamic hand gesture authentication (DHGA) combines physiological and behavioral traits for high-security user verification. However, existing methods suffer from poor generalizability due to data distribution discrepancies across sessions and environments. SSAF addresses this with:
- GE-Stan: A generic preprocessing module that standardizes background, illumination, palm position, angle, and scale.
- AM-Net: A two-stream network decoupling appearance (physiological) and motion (behavioral) features, with an attention-based recoupling mechanism.
- ✅ State-of-the-art accuracy on three DHGA datasets (SCUT-DHGA, SCUT-DHGA-br, SCUT-RealDHGA)
- ✅ 3.6× efficiency boost with minor accuracy trade-off via temporal downsampling
- ✅ Generic preprocessing – plug GE-Stan into most existing DHGA algorithms for significant improvement
- ✅ Real-world dataset – SCUT-RealDHGA with 10 gesture types, diverse backgrounds/illumination, 60 subjects
- ✅ Portable deployment – runs on NVIDIA Xavier edge device with <100ms latency
Standardizes 5 factors causing distribution shifts:
| Factor | Standardization Method |
|---|---|
| Background | Hand segmentation mask |
| Illumination | Brightness normalization |
| Position | Translate palm root to center |
| Angle | Rotate to align center line |
| Scale | Resize to fixed palm length |
- A-stream (Appearance): ResNet18 on sampled RGB frames
- M-stream (Motion): Lightweight A-GCN on skeleton sequences (21 keypoints)
- Recoupling: Cross-attention to inject appearance semantics into motion features
- Adaptive Fusion: Learnable weighted fusion of both streams
| Statistic | Value |
|---|---|
| Subjects | 60 |
| Gesture types | 10 (9 defined + 1 random) |
| Modalities | RGB + Depth |
| Videos | 7,200 |
| Frames | 864,000 (30 fps, 4s each) |
| Scenes | Dormitory, classroom, subway, public hall, etc. |
The SCUT-RealDHGA dataset can be downloaded from SCUT-RealDHGA.
- Python 3.8+
- PyTorch 2.40+
git clone https://github.com/SCUT-BIP-Lab/SSAF.git
cd SSAF
pip install -r requirements.txt- Download datasets: SCUT-DHGA, SCUT-DHGA-br, SCUT-RealDHGA
- Organize as:
data/
├── SCUT-DHGA/
│ ├── color_hand
│ └── keypoint
├── SCUT-DHGA-br/
└── SCUT-RealDHGA/
- Normalize the data using GE‑Stan:
python ./src/utils/GE_Stan.py- The normalized data will be stored as:
data_norm/
├── SCUT-DHGA/
│ ├── color_hand_norm
│ └── keypoint_norm
├── SCUT-DHGA-br/
└── SCUT-RealDHGA/
# Train SSAF on SCUT-DHGA under MG protocol
python ./train.py --conf_file ./conf/SSAF/MG/MG_SD_AMNet.conf --mode train# Evaluate SSAF on SCUT-DHGA under UMG protocol
python ./train.py --conf_file ./conf/SSAF/UMG/UMG1_SD_AMNet.conf --mode evalIf you find this work useful, please cite:
@ARTICLE{zhang2026ssaf,
author={Zhang, Yufeng and Wang, Xilai and Song, Wenwei and Kang, Wenxiong},
journal={IEEE Transactions on Information Forensics and Security},
title={Enhancing Perceptron Constancy for Real-World Dynamic Hand Gesture Authentication},
year={2026},
volume={21},
number={},
pages={886-899},
keywords={Authentication;Hands;Physiology;Accuracy;Skeleton;Videos;Robustness;Feature extraction;Standardization;Lighting;Biometrics;hand gesture authentication;preprocessing;skeleton-based;behavior analysis},
doi={10.1109/TIFS.2025.3648567}}Biometrics and Intelligence Perception Lab
College of Automation Science and Engineering
South China University of Technology, Guangzhou, China
- Yufeng Zhang: auyfzhang@mail.scut.edu.cn
- Wenxiong Kang: auwxkang@scut.edu.cn
MIT License. See LICENSE for details.