Contact: Alexander Hartl, Maximilian Bachl.

This repository contains the code and the figures for the upcoming paper dealing with Explainability and Adversarial Robustness for RNNs. We perform our study in the context of Intrusion Detection Systems.

We use two datasets in the paper: CIC-IDS-2017 and UNSW-NB-15. The repository contains a preprocessed version of the datasets (refer to the paper for more information).

flows.pickle is for CIC-IDS-2017 while flows15.pickle is for UNSW-NB-15. Due to size constraints on GitHub they had to be split. Restore flows.pickle as follows:

• Concatenate the parts: cat flows.pickle.gz.part-* > flows.pickle.gz
• Unzip them: gzip -d flows.pickle.gz

Proceed analogously for flows15.pickle.

If you want to produce the preprocessed files yourself:

• Follow the information on how to reproduce the preprocessed datasets in the Datasets-preprocessing repository.
• Run the parse.py script on the resulting .csv file to get the final .pickle file.

A list of command-line arguments is printed when calling learn.py without arguments. Here we provide a few examples of how learn.py can be called.

# Train a new model
./learn.py --dataroot flows.pickle
# Evaluate a model on the test dataset
./learn.py --dataroot flows.pickle --net runs/Oct26_00-03-50_gpu/lstm_module_1284.pth --function test
# Compute several feature importance metrics
./learn.py --dataroot flows.pickle --net runs/Oct26_00-03-50_gpu/lstm_module_1284.pth --function feature_importance
./learn.py --dataroot flows.pickle --net runs/Nov19_18-25-03_gpu/lstm_module_898.pth --function dropout_feature_importance
./learn.py --dataroot flows.pickle --net runs/Nov19_18-25-03_gpu/lstm_module_898.pth --function dropout_feature_correlation
./learn.py --dataroot flows.pickle --net runs/Oct26_00-03-50_gpu/lstm_module_1284.pth --adjustFeatImpDistribution --function mutinfo_feat_imp
./weight_feat_imp.py runs/Oct26_00-03-50_gpu/lstm_module_1284.pth
# Generate adversarial samples
./learn.py --dataroot flows.pickle --net runs/Oct26_00-03-50_gpu/lstm_module_1284.pth --function adv
# Perform adversarial training
./learn.py --dataroot flows.pickle --net runs/Oct26_00-03-50_gpu/lstm_module_1284.pth --advTraining
# Compute ARS
./learn.py --dataroot flows.pickle --net runs/Oct26_00-03-50_gpu/lstm_module_1284.pth --function adv_until_less_than_half

Models in the runs folder have been trained with the following configurations:

• Oct26_00-03-50_gpu: CIC-IDS-2017
• Oct28_15-41-46_gpu: UNSW-NB-15
• Nov19_18-25-03_gpu: CIC-IDS-2017 with feature dropout
• Nov19_18-25-53_gpu: UNSW-NB-15 with feature dropout
• Nov20_18-27-31_gpu: CIC-IDS-2017 with adversarial training using L1 distance and CW with kappa=1
• Nov20_18-28-17_gpu: UNSW-NB-15 with adversarial training using L1 distance and CW with kappa=1