coldStartPerformance.py

"""
This script checks the performance of the attacks by training the last layer of inception.
The training is cold start
We add the one-poison to the training data and train using the augmented training data
At the end, we check to see whether the target got misclassified or not

developed by ashafahi @ March 13 3:00 pm
"""
import numpy as np
import tensorflow as tf
from util_one_shot_kill_attack import train_last_layer_of_inception
from os import listdir
import imageio

# #load the list directory
# directory = './poisonImages/'
# all_poisons = listdir(directory)
# number = lambda x:int(x[:x.index('_')]) if '.' not in x[:x.index('_')] else 100000
# all_poisons = sorted(all_poisons, key=number)
# #remove all excess files that are not images
# while all_poisons[-1].index('.') == 0:
# 	all_poisons = all_poisons[:-1]

#load the data from file
directorySaving = './Data/XY/'
all_datas = ['X_tr_feats', 'X_tst_feats', 'X_tr_inp', 'X_tst_inp', 'Y_tr', 'Y_tst']
X_tr = np.load(directorySaving+all_datas[0]+'.npy')
X_test = np.load(directorySaving+all_datas[1]+'.npy')
X_inp_tr = np.load(directorySaving+all_datas[2]+'.npy')
X_inp_test = np.load(directorySaving+all_datas[3]+'.npy')
Y_tr = np.load(directorySaving+all_datas[4]+'.npy')
Y_test = np.load(directorySaving+all_datas[5]+'.npy')
print('done loading data!')


Poises = np.load('all_poisons.npy')
other_class_prob = []
numNotMiscclassified = 0
poison_corr_class_prob = []
numPoisonCorr = 0
print(Poises.shape)
for targID, thePoison in enumerate(Poises):
	print("******************%d********************"%targID)
	target_class_probs, target_corr_pred, poison_class_probs, poison_corr_pred = train_last_layer_of_inception(targetFeatRep=X_test[targID],poisonInpImage=thePoison,poisonClass=1-Y_test[targID],X_tr=X_tr,Y_tr=Y_tr,Y_validation=Y_test,X_validation=X_test, cold=True)
	other_class_prob.append(target_class_probs[0][int(1-Y_test[targID])])
	numNotMiscclassified += 1*target_corr_pred[0]
	poison_corr_class_prob.append(poison_class_probs[0][int(1-Y_test[targID])])
	numPoisonCorr += 1*poison_corr_pred[0]
print('##############################')
print('out of %d poisons, %d got correctly classified!'%(len(Poises),numPoisonCorr))
print('out of %d targets, %d got misclassified!'%(len(Y_test),len(Y_test) - numNotMiscclassified))
other_class_prob = np.array(other_class_prob)
np.save('Wrong_ClassProb_targ.npy',other_class_prob)
poison_corr_class_prob = np.array(poison_corr_class_prob)
np.save('Right_ClassProb_poison.npy',poison_corr_class_prob)