yld957 / infoneat Goto Github PK

InfoNEAT is a member of the NEAT (NeuroEvolution of Augmenting Topologies) family. InfoNEAT evolves neural networks and knows when to stop learning and how to scale up as a multi-class algorithm with many (e.g., 256) classes. In a nutshell, two issues with the NEAT algorithm are resolved: application in multi-class scenarios and generic criteria for stopping the training/evolution processes.

This repo presents the InfoNEAT framework used to train NN models that perform SCA. The results will appear in IACR Transactions on Cryptographic Hardware and Embedded Systems (TCHES) (TCHES 2023, Issue 1). It is also available on arxiv.

Python packages to install: scikit-learn, h5py, pickle, csv

Instructions:

1. Extract the files from the folder 'InfoNEAT'.

2. The folder contains a config folder that contains configuration parameters for the NEAT-based algorithm for each dataset. The src folder contains all the source code to train sub-models, stacked models, and even split the dataset into different cross-validation folds.

3. The file 'train.py' contains the code to train sub-models and stacked model for each dataset. Change lines 9-14 accordingly. If needed, set dataset_split = True (line 15) to create cross-folds of the dataset. Please save the .h5 dataset files for each of the datasets under the respective folders.

Result (after running the file 'train.py'):

1. If dataset_split is set to True, then k-folds of the dataset will be created and saved under the datasets folder. (Currently num_of_folds is set to 5)

2. 256 sub-models and a stacked model for a specific fold will be created and saved under the models folder (currently, fold_num is set to 1).

Note: the models/submodels folder contains all the submodels against the three datasets. The models/stacked_models folder contains stacked models for one fold for AES_HD and ASCAD_fixed key datasets. This folder also contains the stacked models for the ASCAD_variable key dataset (no cross-validation was performed for the ASCAD_variable key dataset).

###################################################
Note: modifications to the original neat code can be found in the folder 'neat'. Especially the code to train a submodel using batch sizes of data, training submodels for multiple classes or labels, and the inclusion of CMI criteria to train effective submodel)