Contact: Maximilian Bachl, Alexander Hartl
This repository contains the code, the data, the plots and the tex files for our paper presented at Big-DAMA '19 (arXiv) dealing with backdoors in IDS and ways to defend against them.
Code for the EagerNet paper is in the eager branch.
To train a random forest with the preprocessed UNSW-NB15 dataset, unzip the file CAIA_backdoor_15.csv.gz and run the following command to train a random forest with 100 estimators:
./learn.py --dataroot CAIA_backdoor_15.csv --function train --method rf --backdoor --nEstimators 100
To instead train a neural network, replace --method rf with --method nn.
To reproduce Figure 6 from the paper run:
./learn.py --dataroot CAIA_backdoor_15.csv --method=rf --net <path to you random forest> --function prune_backdoor --backdoor --reduceValidationSet <fraction from 0 to 1 that indicates the fraction of the validation dataset that is going to be used> --nSteps 49 --pruneOnlyHarmless --depth