A novel extreme learning machine-based cryptography system

Extreme Learning Machine (ELM) is a well-known algorithm for the training of neural networks for two modes of functionality: regression and classification. This paper presents a novel model using ELM in ciphering. The study begins with an investigation of the Real-time Recurrent Neural Network (RRNN) derived from the gradient-based learning for symmetric cipher. The weakness of this cipher is that the error converges to zero, and that the RRNN will not change regardless of how the plaintext is changed. Given the nature of the ELM, a technique with an ELM-based cipher is proposed to provide the capability of performing the training independently from the input and the error gradient. Different simulation scenarios were used to evaluate and validate the effectiveness of the proposed cipher. Results revealed that the ELM-based cipher performed better than RRNNs, especially in terms of security. Moreover, the ELM-based cipher demonstrated significantly competing performance for a wide range of evaluation measures. Using an ELM-based cipher instead of an AES or other type of ciphers has the added advantage of providing an addition level of securityby allowing the user to change the algorithmic core of the cipher by simply changing the weights of the neural network. This allows hardware programmed ciphers to be more secure while costing less compared to other ciphers. Copyright (c) 2016 John Wiley & Sons, Ltd.