Reliable Camera Model Identification Using Sparse Gaussian Processes

If you use this implementation, please cite the paper:

@ARTICLE{9392307,
  author={Lorch, Benedikt and Schirrmacher, Franziska and Maier, Anatol and Riess, Christian},
  journal={IEEE Signal Processing Letters}, 
  title={Reliable Camera Model Identification Using Sparse Gaussian Processes}, 
  year={2021},
  volume={28},
  pages={912--916},
  doi={10.1109/LSP.2021.3070206}
}

Prerequisites

1. Set up virtual environment.

conda env create --name gpc --file=environment.yml

2. (Optional) For comparison to the PI-SVM, clone and compile the PI-SVM authors' code from their GitHub repository. Then open utils.constants.py and set constants[LIBSVM_DIR] to the path where to find the compiled executables.

Data preparation

1. Download the Dresden image database.

2. Extract SPAM features.

Example:

PYTHONPATH=`pwd` python spam_features/ddimgdb_extract_features.py \
    --output_dir /media/hdd/camera_model_identification \
    --dresden_dir /mnt/nfs/DresdenImageDB
    [--crop CROP_SIZE]

This script will store the extracted features as HDF5 file to the given output directory. If no crop argument is given, the script extracts the features from the full resolution images.

Train Gaussian process classifier

Example:

PYTHONPATH=`pwd` python experiments/train_spam_gpc.py \
    --dresden_dir /mnt/nfs/DresdenImageDB \
    --features_file $HDD/camera_model_identification/2021_04_13-dresden_spam_features_crop_full_resolution.h5 \
    --logdir $HDD/camera_model_identification/models \
    --num_known_models 10 \
    --max_num_inducing_points 512 \
    --seed 1 \
    --model_selection_seed 91058 \
    --torch_seed 42

In our experiments, we trained each GPC five times with different training-test splits. In particular, we set MODEL_SELECTION_SEED=91058,SEED=$i,TORCH_SEED=42 with i=1,...,5.

Note that there are additional arguments with reasonable default values for controlling the model, the data, and the training.

Evaluation

To evaluate a single trained Gaussian process classifier, you can use eval_single.py.

python experiments/eval_single.py \
    --dresden_dir /mnt/nfs/DresdenImageDB \
    --features_file $HDD/camera_model_identification/2021_04_13-dresden_spam_features_crop_full_resolution.h5 \
    --model_dir $HDD/camera_model_identification/2021_04_14_08_29_42-gpc

To evaluate multiple Gaussian process classifiers trained on the same features, you can use eval_multiple.py.

For comparison against the PI-SVM and the combined classification framework (CCF, named "secure SVM" in the code), you can use eval_and_compare_multiple.py.

Additional experiments provided in this repository

Number of inducing points

1. Train GPC with different numbers of inducing points, see experiments/torque/train_max_num_inducing_points_woody.sh.

2. Evaluate trained GPCs, see experiments/eval_multiple.py.

Impact of ROI size

1. Extract features with different ROI sizes (by running ddimgdb_extract_features.py with the --crop parameter).

2. Train GPC with features of every ROI size, see experiments/torque/train_crops_woody.sh.

3. Evaluate trained GPCs, see experiments/eval_and_compare_multiple.py.

Unseen post-processing

1. With a trained GPC, run experiments/eval_postprocessing.py.