Security and Forensics Exploration of Learning-based Image Coding
Citation
If you find this code useful, please consider citing:
@inproceedings{bhowmik2021vcip,
Author = {Deepayan Bhowmik and Mohamed Elawady and Keiller Nogueira},
Title = {Security and Forensics Exploration of Learning-based Image Coding},
Booktitle = {IEEE Visual Communications and Image Processing (VCIP)},
Year = {2021},
pages={1-5},
doi={10.1109/VCIP53242.2021.9675445}
}
Watermarking
Diagram - Watermarking
Results - Watermarking
Dataset
Kodak Lossless True Color Image Suite this dataset was used to evaluate the watermarking methods.
Usage
You can apply the watermarking process over an image with compression option using the following command exmaple:
python watermarking/main.py --inFolder ./kodak_imgs/ \
--outFolder ./tmp/ \
--imgSrc kodim23.png \
--imgWtr kodim15.png \
--method ADD_DCT \
--comp COMP_TFCI_HI
Source (DNN Architecture) Identification
Diagram - Source Identification
Requirements
Dataset
JPEG-AI Dataset was used to train and evaluate the models.
Compress Images
TensorFlow Compression (TFC) was used to compress images.
After compressing the dataset, its structure should be like this:
dataset/
-training_decoded/
-COMPRESS_METHOD_1/
-img/
-*.png
-COMPRESS_METHOD_2/
-img/
-*.png
-...
-validation_decoded/
-COMPRESS_METHOD_1/
-img/
-*.png
-COMPRESS_METHOD_2/
-img/
-*.png
-...
-test_decoded/
-COMPRESS_METHOD_1/
-img/
-*.png
-COMPRESS_METHOD_2/
-img/
-*.png
-...
Usage
You can train/test a model using the following command:
python src_identification/main.py --operation training/test \
--dataset_path ROOT_PATH_TO_DATASET \
--output_path PATH_TO_SAVE_OUTPUTS_MODELS \
--network NETWORK_TO_USE \
--model_path MODEL_TO_LOAD \
--learning_rate 0.01 \
--weight_decay 0.005 \
--batch_size 128 \
--epoch_num 50