5a3f444696ba821ce35e165747cd03c6e03b6c5a,art/estimators/embedding/keras.py,KerasAdversarialEmbedding,__init__,#KerasAdversarialEmbedding#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,54

Before Change

            detect_threshold=detect_threshold,
        )
        // create discriminator and append to output
        feature_layer_keras = self.model.get_layer(feature_layer)
        dense_layer_1 = Dense(self.discriminator_layer_1)(feature_layer_keras)
        norm_1_layer = BatchNormalization()(dense_layer_1)
        leaky_layer_1 = LeakyReLU()(norm_1_layer)
        dense_layer_2 = Dense(self.discriminator_layer_2)(leaky_layer_1)

After Change

        if not model.losses:
            // Assuming output layer is last layer
            output_layer = len(model.layers) - 1
            losses = {model.layers[output_layer].name: model.loss, "backdoor_detect": "binary_crossentropy"}
        else:
            // TODO: this makes no sense
            losses = {"output_" + str(i + 1): loss for i, loss in model.outputs}

Instances