We are currently trying to use your optimized network architectures for ASCAD with fixed and random key as provided in Tab. 10 and Tab. 12 of the paper Reinforcement Learning for Hyperparameter Tuning in Deep Learning-based Side-channel Analysis .
Unfortunately, the path to the trained model provided in your scripts is not included in the repository.
We tried to build the Keras model for ASCAD random and ID manually
x = keras.layers.Conv1D(128, 3, strides=1, activation='selu', padding='same', kernel_initializer='he_uniform', name='block1_conv1')(x)
x = keras.layers.AveragePooling1D(75, strides=75, name='block1_pool')(x)
x = keras.layers.Flatten(name='flatten')(x)
x = keras.layers.Dense(30, activation='selu', kernel_initializer='he_uniform', name='fc1')(x)
x = keras.layers.Dense(2, activation='selu', kernel_initializer='he_uniform', name='fc2')(x)
x = keras.layers.Dense(256, activation='softmax', kernel_initializer="glorot_uniform", name='predictions')(x)
and use the following code for training
optimizer = keras.optimizers.Adam()
oclr = OneCycleLR(
max_lr=0.005 * 1, end_percentage=0.2, scale_percentage=0.1,
maximum_momentum=None,
minimum_momentum=None, verbose=True
)
self.model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
self.history = self.model.fit(x=self.x_train, y=self.y_train, validation_data=(self.x_val, self.y_val), batch_size=400, verbose=1, shuffle=True, epochs=epochs, callbacks=[callback, best_model_callback, oclr])With these settings, the network is not able to retrieve the key after training and we are wondering whether we misunderstood the settings in the paper.
Could you provide some feedback whether the network model corresponds to the one you proposed in your paper and whether the learning parameters are the same?
Could you also provide the trained networks? This would make it easier to reproduce the results and use your optimized network as a benchmark.
We are currently trying to use your optimized network architectures for ASCAD with fixed and random key as provided in Tab. 10 and Tab. 12 of the paper Reinforcement Learning for Hyperparameter Tuning in Deep Learning-based Side-channel Analysis .
Unfortunately, the path to the trained model provided in your scripts is not included in the repository.
We tried to build the Keras model for ASCAD random and ID manually
and use the following code for training
With these settings, the network is not able to retrieve the key after training and we are wondering whether we misunderstood the settings in the paper.
Could you provide some feedback whether the network model corresponds to the one you proposed in your paper and whether the learning parameters are the same?
Could you also provide the trained networks? This would make it easier to reproduce the results and use your optimized network as a benchmark.