Giaretta, L., Lekssays, A., et al., LiMNet: Early-Stage Detection of IoT Botnets with Lightweight Memory Networks, 2021, in European Symposium on Research in Computer Security (ESORICS 2021)
To run a single run of a single configuration (stored in a configuration file), using GPU number 0 (omit for CPU training) (output stored in runs/limnet-best):
python src/training.py limnet-best src/configs/medbiot-mem-gru-64.py --gpu 0
Example with a configuration file storing several named configurations, each repeated for several runs (output stored in several subfolders with pattern multiruns/all-configs/<config-name>/trial_<N>):
python src/train_all.py all-configs src/configs/all-configs.py --gpu 0
To get a summary of the size of a trained model (in this case using the output of a single configuration run):
python src/model_size.py runs/limnet-best
To measure the inference speed of a model on a single CPU core with batch size of 1 (reasons explained in the paper) (in this case using one particular run for one particular configuration from a multi-configuration, multi-run experiment):
python src/model_speed.py multiruns/all_configs/kitsune-lang-lstm-1-64-all_feats/trial_0
To produce Table 3 or Table 4 from the paper (requires multiruns/all-configs from the second example):
python src/make_latex_tables.py
This code was tested in a Miniconda environment with the following libraries (and their dependencies):
cudatoolkit 11.1 (conda-forge)
ignite 0.5.0.dev20210314 (pytorch-nightly)
numpy 1.19.2
pandas 1.2.3
python 3.8.8
pytorch 1.9.0.dev20210329 (pytorch-nightly)
sklearn 0.24.1
scipy 1.6.2
tqdm 4.56.0