Methods of applying deep learning to database protection have increased over the years. To secure role-based access control (RBAC) by learning the mapping function between query features and roles, it is known that the convolutional neural networks combined with learning classifier systems (LCS) can reach formidable accuracy. However, current methods are focused on using a singular model architecture and fail to fully exploit features that other models are capable of utilizing. Different deep architectures, such as ResNet and Inception, can exploit different spatial correlations within the feature space. In this paper, we propose an ensemble of multiple models with different deep convolutional architectures to improve the overall coverage of features used in role classification. By combining models with heterogeneous topologies, the ensemble-LCS model shows significantly increased performance compared to previous single architecture LCS models and achieves better robustness in the case of training data imbalance.