Deep Learning based Video Watermarking via Group Equivariant Convolutional Networks

Rotation-Invariant Video Watermarking using Group-Equivariant Convolutional Neural Networks

This repository contains the implementation of my Master's thesis work on improving rotation robustness in video watermarking using Group-Equivariant CNNs (G-CNNs). Built upon Meta's VideoSeal framework.

🎯 Key Contribution

Traditional CNN-based watermarking systems struggle with rotation attacks because standard convolutions are not rotation-equivariant. This work introduces a hybrid architecture that combines:

• Standard VideoSeal Embedder — Full expressiveness for watermark embedding
• G-ConvNeXt Extractor — Group-equivariant ConvNeXt for rotation-invariant extraction

Why Hybrid?

Original Image → [Embedder] → Watermarked → [ROTATION ATTACK] → [Extractor] → Message
                     ↑                                              ↑
              Sees ORIGINAL                                  Sees ROTATED
              (no equivariance needed)                       (needs G-CNN!)

Key insight: Rotation happens after embedding, before extraction. Only the extractor needs rotation invariance.

📊 Results

Image Evaluation (COCO - 100 images)

Model	PSNR	Identity	Rotate 30°	Rotate 45°	Rotate 90°
VideoSeal Baseline (22K SA-1B)	44.6 dB	66.1%	51.5%	50.3%	62.9%
G-ConvNeXt C4	44.4 dB	69.6%	61.2%	54.6%	69.4%
G-ConvNeXt C8	44.2 dB	68.4%	59.7%	54.3%	68.1%

Video Evaluation (SA-V - 155 videos)

Model	PSNR	Identity	H264	H265
G-ConvNeXt C4	45.1 dB	67.5%	57.1%	57.7%
G-ConvNeXt C8	43.6 dB	61.7%	54.4%	55.1%

🏗️ Architecture

G-ConvNeXt Extractor

The extractor uses group-equivariant convolutions from the escnn library:

1. Lifting Convolution — Maps input from Z² to group space G × Z²
2. Group Convolutions — Preserves equivariance through the network
3. Group Pooling — Converts equivariant → invariant features for bit prediction

Supported groups:

• C4: 4-fold rotation symmetry (0°, 90°, 180°, 270°)
• C8: 8-fold rotation symmetry (0°, 45°, 90°, ..., 315°)

Parameter Count

Model	Parameters
ConvNeXt-tiny (baseline)	~3.98M
G-ConvNeXt C4	~4.7M
G-ConvNeXt C8	~9.5M

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/sejal-prog/Watermarking.git
cd Watermarking

# Create conda environment
conda create -n gcnn_env python=3.10
conda activate gcnn_env

# Install PyTorch
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.1 -c pytorch -c nvidia

# Install dependencies
pip install -r requirements.txt
pip install escnn  # For group-equivariant convolutions

Training

Image Training (G-ConvNeXt C4)

python train.py \
    --image_dataset sa-1b \
    --video_dataset none \
    --batch_size 8 \
    --gradient_accumulation_steps 2 \
    --img_size 256 \
    --extractor_model g_convnext \
    --embedder_model unet_small2_yuv_quant \
    --nbits 128 \
    --epochs 601 \
    --optimizer AdamW,lr=5e-4 \
    --lambda_dec 1.0 \
    --lambda_d 0.1 \
    --lambda_i 0.1 \
    --perceptual_loss yuv \
    --augmentation_config configs/all_augs.yaml

Note: Set group_type: C4 or group_type: C8 in configs/extractor.yaml

Video Fine-tuning

python train.py \
    --image_dataset sa-1b \
    --video_dataset sav \
    --prop_img_vid 0.0 \
    --video_start 601 \
    --frames_per_clip 16 \
    --batch_size_video 1 \
    --extractor_model g_convnext \
    --embedder_model unet_small2_yuv_quant \
    --nbits 128 \
    --epochs 801 \
    --optimizer AdamW,lr=5e-6 \
    --scaling_w 0.2

Evaluation

python -m videoseal.evals.full \
    --checkpoint /path/to/checkpoint.pth \
    --dataset coco \
    --is_video false \
    --num_samples 100 \
    --output_dir outputs/

📁 Project Structure

├── videoseal/
│   ├── models/
│   │   ├── extractor.py          # Base extractor classes
│   │   └── g_extractor.py        # G-ConvNeXt extractor
│   ├── modules/
│   │   ├── g_convnext.py         # G-ConvNeXt blocks and stages
│   │   └── g_common.py           # Common G-CNN utilities
│   └── evals/
│       └── full.py               # Evaluation script
├── configs/
│   ├── extractor.yaml            # Extractor config (set group_type here)
│   └── all_augs.yaml             # Augmentation config
├── train.py                      # Training script
└── count_params.py               # Parameter counting utility

🔧 Configuration

Setting Group Type

Edit configs/extractor.yaml:

g_convnext:
  group_type: C4  # Options: C4, C8
  encoder:
    depths: [2, 2, 2]
    dims: [32, 64, 128]

📚 References

• VideoSeal: Open and Efficient Video Watermarking
• Group Equivariant Convolutional Networks
• escnn: E(2)-Equivariant Steerable CNNs

📝 Citation

If you use this work, please cite:

@mastersthesis{gcnn_watermarking_2026,
  title={Rotation-Invariant Video Watermarking using Group-Equivariant CNNs},
  author={Sejal},
  year={2026},
  school={Your University}
}

@article{fernandez2024video,
  title={Video Seal: Open and Efficient Video Watermarking},
  author={Fernandez, Pierre and Elsahar, Hady and Yalniz, I. Zeki and Mourachko, Alexandre},
  journal={arXiv preprint arXiv:2412.09492},
  year={2024}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Meta AI Research for the VideoSeal framework
• QUVA Lab for the escnn library
• My thesis supervisors and colleagues