This repository hosts the implementation of our submission to NeurIPS 2023 titled "RS-Del: Edit Distance Robustness Certificates for Sequence Classifiers via Randomized Deletion".
.
βββ configs
β βββ certify-exp # Configs for evaluation step
β βββ models # Configs for malware detection models
β βββ repeat-forward-exp # Configs for sampling step
βββ data
β βββ binaries # Executables for training and evaluation
β βββ {test,train,valid}.csv # CSV files for data partitioning
βββ docker # Docker deployment files
βββ outputs # Directory for experimental outputs
βββ run_scripts # Shell scripts for running experiment steps
βββ src # Source code directory
βββ torchmalware # Python package with core implementations
βββ train.py # Script for training models
βββ repeat_forward_exp.py # Script for sampling perturbed inputs
βββ fp_curve-repeat_forward.py # Script for computing FPR curve
βββ certify_exp-repeat_forward.py # Script for computing certified radius
- Train the smoothed model using data augmentation via
src/train.py
. - Example: See
run_scripts/task1-train.sh
.
python src/train.py --conf configs/models/sample_config.yaml
- Save base model confidence scores via
src/repeat_forward_exp.py
. - Example: See
run_scripts/task2-repeat_forward.sh
.
python src/repeat_forward_exp.py --conf configs/repeat-forward-exp/sample_config.yaml
- Vary the decision threshold and compute the FPR via
src/fp_curve-repeat_forward.py
. - Example: See
run_scripts/task3-fp_curve.sh
.
python src/fp_curve-repeat_forward.py --path model/checkpoint.pth --repeat-conf configs/repeat-forward-exp/sample_config.yaml
- Compute the certified radius via
src/certify_exp-repeat_forward.py
. - Example: See
run_scripts/task4-certify-repeat_forward.sh
.
python src/certify_exp-repeat_forward.py --repeat-conf configs/repeat-forward-exp/sample_config.yaml --certify-conf configs/certify-exp/sample_config.yaml
Execute the steps in the provided Docker container.
git clone $REPO_NAME $DEST
cd $DEST/run_scripts
chmod +x ./run.sh
./run.sh -p $SH_PATH -m $MEM -c $NUM_CORES -g $GPU_ID
- For sequential execution of all steps (1-4), use
run_scripts/task-full.sh
(Not recommended due to long running time).
For reproducing experiments on your dataset, follow the instructions in data/README.md
.
This project is licensed under the MIT License - see the LICENSE.md file for details.
@inproceedings{huang2023rsdel,
author = {Huang, Zhuoqun and Marchant, Neil and Lucas, Keane and Bauer, Lujo and Ohrimenko, Olya and Rubinstein, Benjamin I. P.},
title = {{RS-Del}: Edit Distance Robustness Certificates for Sequence Classifiers via Randomized Deletion},
year = {2023},
booktitle = {Advances in Neural Information Processing Systems},
series = {NeurIPS},
}