==============================

Optimized hyperparameters can be found for the non-contrastive SSL models in best_config.yml, and the optimized hyperparameters for the two baselines are stored in best_config_baselines.yml

Installation

git clone https://github.com/renje4z335jh4/non_contrastive_SSL_NIDS.git
conda create --name [env_name] python=3.8
conda activate [env_name]
cd non_contrastive_SSL_NIDS
pip install -e .

Usage Guide

Datasets

5G-NIDD

You can download the Combined.zip here

Please extract the csv-file in 'data/raw/5GNIDD/'

UNSW-NB15 (If the links get expired, check the official website of the dataset)

UNSW_NB15_Test.csv: here

UNSW_NB15_Train.csv: here

Please extract the csv-file in 'data/raw/UNSW-NB15/'

Processing the Datasets

The process script can be used as follows:

python src/data/process.py [data_set] -d [path/to/dir/containing/the/CSV/files] -o [path/to/output/dir]

Example

python src/data/process.py UNSW-NB15 -d data/raw/UNSW-NB15/ -o data/processed/

Reproduce the Experiments

Hyperparameter tuning

To reproduce the experiments of the paper, we will start by executing the hyper-optimization. This can be simply done by executing:

src/hyperopt/full_hyperopt.sh

Details

The bash script will run three python files, which are hyper-optimizing each combination of model, encoder and augmentation on the data sets UNSW-NB15 and 5G-NIDD, where the models are BarlowTwins, BYOL, SimSiam, VICReg and W-MSE; the encoders are a CNN, a MLP and a FT-Transformer; and the augmentations are Gaussian Noise, Mixup, Random Shuffle, Subsets, SwapNoise (also called CutMix) and Zero Out Noise. In total 5 * 6 * 3 = 90 different experiments are tuned for the two datasets.

Step by step the script is optimizing:

1. Model and augmentation parameters
2. Learning rate
3. Number of epochs to train the model

Similarly, the baselines Deep AutoEncoder and DeepSVDD are tuned with

src/hyperopt/full_hyperopt_baselines.sh

Note: This could take a significant amount of time and resources!

Model Training

The hyper-parameters collected in the tuning step are stored in a YAML file. The final results are gathered by training the different models with the best hyper-parameters for 10 runs and averaging the performance metrics. This is done by executing

python src/utils/run_models.py

The result of each model together with the weights of the best performing run (depicted by the AUROC) is stored under run_results/.

In a similar way, results of the optimized baselines are generated by executing

python src/utils/run_baselines.py

Useful information about the trainer parameter

1. Parameters like model, batch_size, n_epochs, device are self-explaining
2. anomaly_label parameter indicates which label is the malicious label (commonly 1)
3. ckpt_root - if given a checkpoint, will be saved every 5 epochs. Set to None, if you only want to save the model regularly
4. safe_best_model - if given the trainer, will save the best model (lowest loss) as checkpoint.