def fit_discriminator(self,
iterations,
batch_size):
if self.discrimination_mode == DiscriminatorGame.DetectEncryption:
msgs, keys = gen_symmetric_encryption_data(iterations *
batch_size,
self.message_length)
results = self.encryptor.predict([msgs, keys])
P, C, Y = replace_encryptions_with_random_entries(msgs,
results,
real_label=self.discriminator_real_label)
return self.discriminator.fit(
x=[P, C],
y=Y,
epochs=1,
batch_size=batch_size,
verbose=self.verbose
).history
else:
raise NotImplementedError('this GAN mode is not yet supported')
def fit_generator(self, iterations, batch_size):
if self.discrimination_mode == DiscriminatorGame.DetectEncryption:
msgs, keys = gen_symmetric_encryption_data(iterations * batch_size,
self.message_length)
real_labels = np.zeros(
len(msgs)) if self.discriminator_real_label == 1 else np.ones(
len(msgs))
return self.generator.fit(x=[msgs, keys],
y=[msgs, real_labels],
epochs=1,
batch_size=batch_size,
verbose=self.verbose).history
else:
raise NotImplementedError('this GAN mode is not yet supported')
def __call__(self,
epochs=50,
bob_prefit_epochs=10,
eve_postfit_epochs=50,
eve_minibatch_multiplier=2,
iterations_per_epoch=100,
batch_size=512):
"""
fit the model around randomly generated message/key pairs.
:param epochs: the number of epochs to adversarially train for
:param bob_prefit_epochs: the number of epochs to pre-fit bob for
:param eve_postfit_epochs: the number of epochs to post-fit bob for
:param eve_minibatch_multiplier: the minibatch multiplier for eve
:param iterations_per_epoch: the iterations per epoch
:param batch_size: the batch size
:return: a dictionary of agent keys and concatenated training history values.
when an agent is not fit while another is, their history is populated with
float('nan') values.
"""
bob_histories = []
eve_histories = []
for i in range(0, bob_prefit_epochs):
msgs, keys = gen_symmetric_encryption_data(iterations_per_epoch *
batch_size)
self.print('pre-fitting bob')
bob_history = self.bob.fit(x=[msgs, keys],
y=[msgs, msgs],
epochs=1,
batch_size=batch_size,
verbose=self.verbose).history
bob_histories.append(bob_history)
eve_histories.append(nanify_dict_of_lists(bob_history))
self.save()
for i in range(0, epochs):
msgs, keys = gen_symmetric_encryption_data(iterations_per_epoch *
batch_size)
self.print('fitting bob')
bob_history = self.bob.fit(x=[msgs, keys],
y=[msgs, msgs],
epochs=1,
batch_size=batch_size,
verbose=self.verbose).history
bob_histories.append(bob_history)
msgs, keys = gen_symmetric_encryption_data(
eve_minibatch_multiplier * iterations_per_epoch * batch_size)
self.print('fitting eve')
eve_history = self.eve.fit(x=[msgs, keys],
y=[msgs],
epochs=1,
batch_size=2 * batch_size,
verbose=self.verbose).history
eve_histories.append(eve_history)
self.save()
for i in range(0, eve_postfit_epochs):
msgs, keys = gen_symmetric_encryption_data(
iterations_per_epoch * eve_minibatch_multiplier * batch_size)
print('fitting eve')
eve_history = self.eve.fit(x=[msgs, keys],
y=[msgs],
epochs=1,
batch_size=2 * batch_size,
verbose=self.verbose).history
eve_histories.append(eve_history)
bob_histories.append(nanify_dict_of_lists(eve_history))
self.save()
history = self.generate_cohesive_history({
'alice / bob':
join_list_valued_dictionaries(*bob_histories),
'eve':
join_list_valued_dictionaries(*eve_histories)
})
self.history = history
return history