Read Keras model. Read deployment data, which are normalized into range [0,1]. Apply the Keras model to the deployment data, calculate reconstruction errors. Based on the reconstruction errors and a predefined threshold, assign label "fraud"/"normal" to the transactions in the deployment data.