This repo contains a the code for TrinitySRITrojAI submission to the cyber-pdf-dec2022 round of the TrojAI leaderboard.

Dependencies

First install the dependencies required by https://github.com/usnistgov/trojai-example.

Additional dependencies

pip install pandas
pip install scipy
pip install scikit-learn

Usage

Clone code into <root>/trojai-cyber-pdf

Download and extract training set cyber-pdf-dec2022-train.tar.gz into <root>/trojai-datasets/cyber-pdf-dec2022-train

cd into <root>/trojai-cyber-pdf and run the following commands

First run feature extraction on the training set.

python weight_analysis.py

This will produce a feature file <root>/trojai-cyber-pdf/data_cyber-pdf_weight.pt.

Then run cross-validation hyperparameter search using the feature file and a pre-defined detector architecture

python crossval_folds_v2.py --arch arch.mlp_set5 --data data_cyber-pdf_weight.pt --nsplits 7

This will produce a set of learned detector parameters at <root>/trojai-cyber-pdf/sessions/0000000/.

Finally copy the detector parameters into a learned_parameters folder and build the singularity container.

cp -r ./sessions/0000000/ ./learned_parameters
python metafile_generator.py
./build.sh

The script will test inference functionalities and build a container at cyber-pdf-dec2022_sts_SRI_weight_grad_v1.simg.