def from_file(filename):
"""
Loads a DDWitness object from a file with given name.
"""
log.debug("Loading result from file '%s'", filename)
with open(filename, "rb") as f:
obj = pickle.load(f)
assert (type(obj) == DDWitness)
return obj
def _transform(self, x):
"""
Perform feature transformation.
"""
if self.feature_transform is not None:
log.debug("Transforming features...")
x = self.feature_transform.transform(x)
log.debug("Done transforming features")
return x
def _get_tensorboard_log_dir(self):
now = datetime.now()
t = now.strftime("%Y-%m-%d_%H-%M-%S")
letters = string.ascii_letters
rand = ''.join(random.choice(letters) for _ in range(10))
details = [t, rand, self.label]
details = '_'.join(str(d) for d in details)
d = os.path.join(self.tensorboard_dir, details)
log.debug('Tensorboard directory at ' + d)
return d
def to_tmp_file(self) -> str:
"""
Stores the result to a temporary file.
Returns:
The path of the created temporary file.
"""
tmp_dir = get_output_directory("tmp")
fd, filename = tempfile.mkstemp(dir=tmp_dir)
log.debug("Storing result to file '%s'", filename)
with os.fdopen(fd, "wb") as f:
pickle.dump(self, f)
return filename
def best_attack(self, a1, a2) -> MlAttack:
"""
Runs the optimizer to construct an attack for given input pair a1, a2.
Args:
a1: 1d array representing the first input
a2: 1d array representing the second input
Returns:
The constructed MlAttack
"""
log.debug("Searching best attack for mechanism %s, classifier %s...",
type(self.mechanism).__name__,
type(self.classifier_factory).__name__)
classifier = self._train_classifier(a1, a2)
with time_measure("time_determine_threshold"):
log.debug("Determining threshold...")
# TODO: maybe parallelize this loop?
probabilities = []
for parallel_size in split_into_parts(self.config.n,
self.config.n_processes):
sequential_probabilities = []
for sequential_size in split_by_batch_size(
parallel_size, self.config.prediction_batch_size):
# generate samples from a2
b_new = self.mechanism.m(a2, sequential_size)
if len(b_new.shape) == 1:
# make sure b1 and b2 have shape (n_samples, 1)
b_new = np.atleast_2d(b_new).T
# compute Pr[a1 | M(a1) = b_new]
probabilities_new = classifier.predict_probabilities(b_new)
# wrap up
sequential_probabilities.append(probabilities_new)
sequential_probabilities = np.concatenate(
sequential_probabilities)
probabilities.append(sequential_probabilities)
probabilities = np.concatenate(probabilities)
probabilities[::-1].sort() # sorts descending, in-place
assert (probabilities.shape[0] == self.config.n)
# find optimal threshold
log.debug("Finding optimal threshold...")
with time_measure("time_dp_distinguisher_find_threshold"):
thresh, q = DPSniper._find_threshold(
probabilities, self.config.c * probabilities.shape[0])
log.debug("Selected t = %f, q = %f", thresh, q)
return MlAttack(classifier, thresh, q)
def _train_one_batch(self, batch_idx: int, x_train, y_train, x_test,
y_test, writer, optimizer, scheduler):
for epoch in range(self.epochs):
# not really an "epoch" as it does not loop over the whole data set, but only over one batch
# closure function required for optimizers such as LBFGS that need to compute
# the gradient of the loss themselves
def closure():
# initialize gradients
optimizer.zero_grad()
# compute loss (forward pass)
y_pred = self.model(x_train).squeeze()
loss = self._regularized(self.criterion(y_pred, y_train))
# backward pass
loss.backward()
return loss
optimizer.step(closure)
if scheduler is not None:
scheduler.step()
# recompute loss for logging
loss = closure()
# logging
step_idx = (batch_idx - 1) * self.epochs + epoch
writer.add_scalar('Loss/train', loss.item(), step_idx)
log.debug(
'Batch {}, epoch {} (global step {}): train loss: {}'.format(
batch_idx, epoch, step_idx, loss.item()))
if x_test is not None:
# compute test loss
y_pred_test = self.model(x_test).squeeze()
loss_test = self.criterion(y_pred_test, y_test.squeeze())
writer.add_scalar('Loss/test', loss_test.item(), step_idx)
# compute testing accuracy
accuracy_test = 1 - torch.mean(
torch.abs(torch.round(y_pred_test) - y_test))
writer.add_scalar('Accuracy/test', accuracy_test.item(),
step_idx)
def _generate_data_batch(self, a1, a2, n) -> Tuple:
"""
Generates a training data batch of size 2n (n samples for each input a1 and a2).
"""
log.debug("Generating training data batch of size 2*%d...", n)
b1 = self.mechanism.m(a1, n)
b2 = self.mechanism.m(a2, n)
if len(b1.shape) == 1:
# make sure b1 and b2 have shape (n_samples, 1)
b1 = np.atleast_2d(b1).T
b2 = np.atleast_2d(b2).T
# rows = samples, columns = features
x = np.concatenate((b1, b2), axis=0)
# 1d array of labels
y = np.zeros(2 * n)
y[n:] = 1
return x, y
def compute_eps_estimate(self, a1, a2, attack: Attack) -> (float, float):
"""
Estimates eps(a2, a2, attack) using samples.
Returns:
a tuple (eps, lcb), where eps is the eps estimate and lcb is a lower confidence bound for eps
"""
p1 = self.pr_estimator.compute_pr_estimate(a1, attack)
p2 = self.pr_estimator.compute_pr_estimate(a2, attack)
log.debug("p1=%f, p2=%f", p1, p2)
log.data("p1", p1)
log.data("p2", p2)
if p1 < p2:
if self.allow_swap:
p1, p2 = p2, p1
log.debug("swapped probabilitites p1, p2")
else:
log.warning("probability p1 < p2 for eps estimation")
eps = self._compute_eps(p1, p2)
lcb = self._compute_lcb(p1, p2)
return eps, lcb
def _compute_results_for_all_inputs(self):
log.debug("generating inputs...")
inputs = []
for (a1, a2) in self.input_generator.get_input_pairs():
log.debug("%s, %s", a1, a2)
inputs.append((self, a1, a2))
log.debug("submitting parallel tasks...")
result_files = the_parallel_executor.execute(DDSearch._one_input_pair,
inputs)
log.debug("parallel tasks done!")
results = []
for filename in result_files:
cur = DDWitness.from_file(filename)
os.remove(filename)
results.append(cur)
return results
def _train_classifier(self, a1, a2) -> StableClassifier:
"""
Trains the classifier for inputs a1, a2.
"""
def generate_batches():
for size in split_by_batch_size(self.config.n_train,
self.config.training_batch_size):
yield self._generate_data_batch(a1, a2, size)
log.debug("Creating classifier...")
classifier = self.classifier_factory.create()
log.debug("Training classifier...")
with time_measure("time_dp_distinguisher_train"):
classifier.train(generate_batches())
log.debug("Done training")
return classifier
def run(self):
log.info("using configuration %s", self.config)
attack_opt = DPSniper(self.mechanism, self.classifier_factory, self.config)
with time_measure("time_dd_search"):
log.debug("running dd-search...")
opt = DDSearch(self.mechanism, attack_opt, self.input_pair_sampler, self.config)
res = opt.run()
log.debug("finished dd-search, preliminary eps=%f", res.eps)
with time_measure("time_final_estimate_eps"):
log.debug("computing final eps estimate...")
res.compute_eps_high_precision(self.mechanism, self.config)
log.info("done!")
log.info("> a1 = {}".format(res.a1))
log.info("> a2 = {}".format(res.a2))
log.info("> attack = {}".format(res.attack))
log.info("> eps = {}".format(res.eps))
log.info("> eps lcb = {}".format(res.lower_bound))
log.data("eps", res.eps)
log.data("eps_lcb", res.lower_bound)
def _one_input_pair(task):
optimizer, a1, a2 = task
pr_estimator = EpsEstimator(optimizer.pr_estimator)
log.debug("selecting attack...")
with time_measure("time_dp_distinguisher"):
attack = optimizer.attack_optimizer.best_attack(a1, a2)
log.debug("best attack: %s", attack)
cur = DDWitness(a1, a2, attack)
log.debug("computing estimate for eps...")
with time_measure("time_estimate_eps"):
cur.compute_eps_using_estimator(pr_estimator)
log.debug("current eps: %s", cur.eps)
log.data("eps_for_sample", cur.eps)
log.debug("storing result...")
filename = cur.to_tmp_file()
log.debug("done!")
return filename
