Parallel Implementation of Chaos Neural Networks for an Embedded GPU

The Internet of Things (IoT) has become ubiquitous, and the need for higher information security is increasing. The CPU usage cost of IoT devices to process information security tasks is large. In the present paper, we study a parallel implementation of chaos neural networks for an embedded GPU using the Open Computing Language (OpenCL). We evaluate this parallel implementation, and the results indicate that it can extract a pseudo-random number series at high speed and with low CPU usage. This implementation is remarkably faster than the implementation in the CPU and is approximately 49% faster than AES in counter mode. The rate of pseudo-random number generation is higher than 2.1 Gbps when using 100 compute units of a GPU. Applying a stream cipher is sufficient even for Internet communication. Extracted pseudo-random number series are independent, have fine randomness properties, and can merge into one series applied to a stream cipher.