ANN version of EACirc
The main scripts are eanet_ann.py for basic aritfitial neural network and eanet_cnn.py for convoluted version, where the first layer groups input into bytes.
Both scripts has same interface (and almost same code (it could be nice to merge them together with one more argument).
> python eanet_ann.py --in_a REFERENCE.bin --in_b TESTED.bin --tv_size BYTE_SIZE_OF_TEST_VECTOR
Where TESTED.bin is binary file for test, REFERENCE.bin is binary file with truly random data (or strong PRNG). The BYTE_SIZE_OF_TEST_VECTOR sets the width of first layer and it should correspond to the source data (AES has 16 B block, so set at least 16 B).
The output is:
> Using TensorFlow backend.
> Preparing test vectors
> Training started
> Epoch 1/100
> 2500/2500 [==============================] - 0s - loss: 0.6973 - acc: 0.5084
...
> Epoch 100/100
> 2500/2500 [==============================] - 0s - loss: 0.0372 - acc: 1.0000
> Trained
> Preparing final test vectors
> 12500/12500 [==============================] - 0s
> Score = [0.042080752551555634, 0.99936002492904663]
Simplified, the acc means what ratio of correct guesses hat the ANN on the data during given batch. The loss says, how fast the ANN learns - lower numbers imply smaller changes - fine grained fine-tuning of the values.
The final evaluation is done on fresh data (and much bigger batch). The score is tuple of loss and acc. The second value for random data should act accordingly to binomial distribution with n=12500 and p=0.5. The p-value of this computation would be 1. Even acc=0.55 is significant for that big batch.
The script run_ann.py enables running scripts eanet_ann.py and eanet_cnn.py on multiple binary files in sequence. It expects you have reference file named in_b.bin for running ANN scripts. Then it runs given script over all binary files in current directory. If the binary files have naming convention NAME_bSIZE.bin for example AES_r3_b16.bin, it will pass the corresponding block size to the ANN script (in the example, it will set --tv_size 16). If the size is not set, the default value is 16 B.
> python3 run_ann.py --script SCRIPT_NAME
The output is just persistent keras info, that we are using TensorFlow backend. But the script creates SCRIPT_NAME_result.json result file and also loggs the computation to ann_runner.log.
The eanet_ann.py_result.json can be processed by process.py script, that creates figures of the learning process of the ANN computations and highlights the final values. It has no arguments, but the processed files have to be named result_ann.json and result_cnn.json. You also need to have directories: ann and cnn, where the figures are stored.