Abstract
The homomorphic cryptographic operations is an umbrella term for computation performed on encrypted data without explicit decryption. The purpose of these operations is to manipulate encrypted data without having to apply decryption first and therefore minimize the computational overhead, breach of anonymity, privacy and without having to disclose private content. One of the promising prospects of homomorphic cryptography is the data classification using neural networks mounted to the back-planes of computational clouds or IoT sensors. While several approaches already explored the classification of encrypted data on specific ciphers, it is yet not well known how well such tasks can be performed on the state of the art AES encryption which was never designed to be homomorphic. In order to provide some insight on this topic, we investigate three different aspects of the classification of AES encrypted data: end-to-end learning, transfer learning and the ability of learning the cipher in the context of classification. We compare the performance of network models trained using transfer learning with end-to-end trained models on encrypted data. We also evaluate the classification of encrypted images using Invertible Neural Network (INN) as a mean to learn and predict the encryption of the data, as well as to determine if the learned AES can be efficiently learned. Finally using INN, we evaluate the learning and memorization extent of the encryption: we perform cross-data validation on different combinations of MNIST datasets such as handwritten digits, fashion images and handwritten letters.