- replicates an exisiting method that utilize SimCLR on intrusion detection
- replace the backbone network of the extractor to MLP
In the files of experiment-<model-name>.ipynb, we hyperparameter searched the best params to be evaluate the performance
In the files of train-<model-name>.ipynb, we trained the models with best params and saved the preprocessor and the models along with 5-step-wise checkpoints.
In the files of eval.ipynb, we evaluate the resnet based extractor(original) and the MLP based extractor(ours) on a set of classic ML classifiers.
All the other files are clearly named the intent so no more introduction to those files.
We found
- It’s possible to use an extractor with a simpler backbone network to substitute.
- The extractor with a simpler backbone network perform similar when on non-linear learning methods and not- learning-based methods while perform worse on linear learning methods.
- It significantly(^10) reduces the computation resources needed by a feature extractor.