def get(self):
data = self.parser.parse_args()
img = data.get('img')
img_data = img.split(',')
img = np.array(img_data, np.float32).reshape(28, 28)
img = img * 255.0
img_new = np.zeros((1, 32, 32, 1))
img_new[0] = np.pad(img.reshape(28, 28), [(2, ), (2, )],
mode='constant').reshape(32, 32, 1)
global sess
global graph
with graph.as_default():
set_session(sess)
attack = HopSkipJump(classifier=classifier,
targeted=False,
max_iter=0,
max_eval=1000,
init_eval=10)
iter_step = 3
x_adv = None
for i in range(iter_step):
x_adv = attack.generate(x=img_new,
x_adv_init=x_adv,
resume=True)
#clear_output()
# print("Adversarial image at step %d." % (i * iter_step),
# "and class label %d." % np.argmax(classifier.predict(x_adv)[0]))
attack.max_iter = iter_step
sav_img = Image.fromarray(x_adv.reshape(32, 32))
sav_img = sav_img.convert("L")
sav_img.save("test.jpg")
buffer = BytesIO()
sav_img.save(buffer, format="JPEG")
myimage = buffer.getvalue()
res = str(predict(x_adv))
print("After Attack: ", res)
return jsonify({
'res': res,
'dat': bytes.decode(base64.b64encode(myimage))
})
attack = HopSkipJump(classifier=classifier,
targeted=True,
max_iter=0,
max_eval=1000,
init_eval=10)
iter_step = 10
stop = Image.open(curr_path + "../danny-machine/machine.jpg")
stop = np.array([np.array(stop)]).astype(float)
x_adv = stop
errors = []
for i in range(100):
x_adv = attack.generate(x=np.array([target_image]),
y=[1],
x_adv_init=x_adv)
l2_err = np.linalg.norm(np.reshape(x_adv[0] - target_image, [-1]))
print("Adversarial image at step %d." % (i * iter_step), "L2 error",
np.linalg.norm(np.reshape(x_adv[0] - target_image, [-1])),
"and class label %d." % np.argmax(classifier.predict(x_adv)[0]))
errors.append((i * iter_step, l2_err))
im = Image.fromarray(np.reshape(x_adv[0].astype(np.uint8), SHAPE))
im.save(curr_path + f"../danny-machine/step{i}.png")
#plt.imshow(np.reshape(x_adv[0].astype(np.float32), (400, 400)))
#plt.show(block=False)
attack.max_iter = iter_step
print(errors)
open('errs.txt', 'r').write(str(errors))