def corrtest_trace_set(trace_set, result, conf=None, params=None):
logger.info("corrtest %s" % (str(params) if not params is None else ""))
if trace_set.windowed:
# Get params
if params is None:
model_type = "aicorrnet" # TODO model_type can be inferred from conf. Therefore change AI to only require conf.
else:
model_type = str(params[0])
if result.ai is None:
result.ai = ai.AI(conf, model_type)
result.ai.load()
# Fetch inputs from trace_set
x = AIInput(conf).get_trace_set_inputs(trace_set)
# Get encodings of signals
encodings = result.ai.predict(x)
# Replace original signal with encoding
assert (encodings.shape[0] == len(trace_set.traces))
for i in range(0, len(trace_set.traces)):
trace_set.traces[i].signal = encodings[i]
# Adjust window size
trace_set.window = Window(begin=0, end=encodings.shape[1])
trace_set.windowed = True
else:
logger.error(
"The trace set must be windowed before testing can take place because a fixed-size input tensor is required by Tensorflow."
)
def _preprocess_trace_set(self, trace_set):
"""
Preprocess trace_set specifically for AutoEncoder
"""
# Get model inputs (usually the trace signal)
signals = AIInput(self.conf).get_trace_set_inputs(trace_set)
return signals, signals # Model output is same as model input
def _preprocess_trace_set(self, trace_set):
"""
Preprocess trace_set specifically for AICorrNet
"""
# Get model inputs (usually the trace signal)
signals = AIInput(self.conf).get_trace_set_inputs(trace_set)
# Get model labels (key byte leakage values to correlate / analyze)
values = LeakageModel(self.conf).get_trace_set_leakages(trace_set)
return signals, values
def on_epoch_end(self, epoch, logs=None):
logs = logs or {}
if epoch % self.metric_freq != 0 or epoch == 0:
return
if self.trace_set is not None:
# Fetch inputs from trace_set
x = AIInput(self.conf).get_trace_set_inputs(self.trace_set)
if self.cnn:
x = np.expand_dims(x, axis=-1)
encodings = self.model.predict(x) # Output: [?, 16]
# Store encodings as fake traceset
keys = np.array([trace.key for trace in self.trace_set.traces])
plaintexts = np.array(
[trace.plaintext for trace in self.trace_set.traces])
fake_ts = TraceSet(traces=encodings,
plaintexts=plaintexts,
keys=keys,
name="fake_ts")
fake_ts.window = Window(begin=0, end=encodings.shape[1])
fake_ts.windowed = True
for i in range(self.key_low, self.key_high):
if len(set(keys[:, i])) > 1:
print(
"Warning: nonidentical key bytes detected. Skipping rank calculation"
)
print("Subkey %d:" % i)
print(keys[:, i])
break
rank, confidence = calculate_traceset_rank(
fake_ts, i, keys[0][i], self.conf
) # TODO: It is assumed here that all true keys of the test set are the same
self._save_best_rank_model(rank, confidence)
logs['rank %d' % i] = rank
logs['confidence %d' % i] = confidence
#self._save_best_rank_model(np.mean(ranks))
else:
print("Warning: no trace_set supplied to RankCallback")
def salvis(self, trace_set_paths, model_type, vis_type, conf):
"""
Visualize the salience of an AI.
:param self:
:param trace_set_paths: List of trace set paths to be used as possible examples for the saliency visualization.
:param model_type: Type of model to load for this configuration.
:param conf: Configuration of the model (required preprocessing actions, architecture, etc.).
:return:
"""
logger.info("Loading model")
model = ai.AI(conf, model_type)
model.load()
logger.info("Resolving traces")
resolve_paths(trace_set_paths)
examples_iterator, _ = aiiterators.get_iterators_for_model(
model_type,
trace_set_paths, [],
conf,
hamming=conf.hamming,
subtype=None,
request_id=self.request.id)
logger.info("Retrieving batch of examples")
trace_set = examples_iterator.get_all_as_trace_set(
limit=int(conf.saliency_num_traces / 256))
examples_batch = AIInput(conf).get_trace_set_inputs(trace_set)
examples_batch = examples_batch[0:conf.saliency_num_traces, :]
if len(examples_batch.shape) != 2:
raise ValueError(
"Expected 2D examples batch for saliency visualization.")
if conf.saliency_remove_bias:
examples_batch = examples_batch[:, 1:]
kerasvis = True if vis_type == 'kerasvis' else False
return SalvisResult(examples_batch=examples_batch,
gradients=saliency.get_gradients(conf,
model,
examples_batch,
kerasvis=kerasvis))