MLSecurityProject

this is the final project code for MLS 2021

For the first method:

Since the second method has better performance, the eval program of the first method would only return the accuracy and asr on the validation data, does not support the predict for a single image. usage:

python mask.py model_filename clean_data_filename poison_data_filename
# output: clean accuracy and asr

example:

python mask.py models/bd_net.h5 data/cl/valid.h5 data/bd/valid.h5
# output: 
# Clean Classification accuracy: 79.44054732831039
# Attack Success Rate: 8.539014462630986

For the second method:

usage:

python p_eval.py model_filename data_filename img_filename
# output: a number

example:

python p_eval.py models/bd_net.h5 data/valid.h5 img.png
# output: 1283