facenet3

import matplotlib.pyplot as plt
import copy

# get session from keras to use it with tf
sess = K.get_session()

# keep track of l2 distance between the generated image and target image.
dist = []

# this will hold the generated image
new_img = copy.deepcopy(attacker)

noise = None
for i in range(140):
    # we pass target and attacker images as mentioned earlier
    gradout = sess.run(grad[0], feed_dict={"input_1:0":np.array([new_img, target, attacker]), K.learning_phase(): 0})

    # if the max gradient is too small, we set a small value
    grad_absmax = max(np.abs(gradout[0]).max(), 1e-10)

    # by doing this way, we get a faster convergence while gradients don't vanish soon.
    step_size = 0.0003 / grad_absmax

    # update noise
    if noise is None:
        noise = step_size * gradout[0]
    else:
        noise -= step_size * gradout[0]

    # also note that, one of the goals is to not visibly alter the image. So we limit the changes to noise.
    noise = np.clip(a=noise, a_min=-0.012,
                           a_max=0.012)

    # apply noise to the attacker image
    new_img = copy.deepcopy(attacker) + noise
    # if the values go beyond 1 or fall below 0, the image will be invalid. So clip it to keep it in range.
    new_img = np.clip(a=new_img, a_min=0.0, a_max=1.0)

    # l2 distance between embeddings of target image and the generated image
    dist.append(np.linalg.norm(database["kian"] - FRmodel.predict_on_batch(np.array([new_img]))))

plt.plot(dist)