Deep-Learning-Based SDN Model for Internet of Things: An Incremental Tensor Train Approach

The Internet of Things (IoT) has emerged as a revolution for the design of smart applications like intelligent transportation systems, smart grid, healthcare 4.0, Industry 4.0, and many more. These smart applications are dependent on the faster delivery of data which can be used to extract their inherent patterns for further decision making. However, the enormous data generated by IoT devices are sufficient to choke the entire underlying network infrastructure. Most of the data attributes present little or no relevance to the prospective relationships and associations with the projected benefits foreseen. Therefore, order-based generalization mechanisms, known as tensors, can be used to represent these multidimensional data, thereby minimizing the flow table (FT) lookup time and reducing the storage occupancy. So, a novel IoT-train-deep approach for intelligent software-defined networking is designed in this article. The proposed approach works in four phases: 1) tensor representation; 2) deep Boltzmann machine-based classification; 3) subtensor-based flow matching process; and 4) incremental tensor train network for FT synchronization. The proposed model has been extensively tested, and it illustrates significant improvements with respect to delay, throughput, storage space, and accuracy.

I. Introduction

Smart cities generate huge amounts of data from numerous connected devices (sensors, vehicles, actuators, etc) and applications (traffic management, healthcare system, stock exchange data, e-shopping, etc.) which are transmitted to geolocated sites for processing or storage [1]. For example, in the healthcare sector, a large amount of medical images and data are generated and stored for use in online medical health records. Similarly, at the consumer end, the content (images, videos, etc.) is generated via social networks for creating a highly connected society. Another example is that of smart meters, deployed at each household, capture the energy usage data continuously thereby generating a multithousand fold data. These kinds of multihundred or multithousand petabyte data sets are creating a new horizon of opportunities in diverse application domains like human genomics, healthcare systems, finance and banking, oil and natural gas exploration, and many more. These data being huge in amount (volume), variable in type and structure (velocity), and generated at different rates or times (variety) have to be analyzed using sophisticated algorithms or big data solutions for identifying hidden patterns which can further lead to several benefits related to decision making in smart ecosystems [2]. In simple words, big data technologies and architectures are designed to extract economical value from a large amount of variable data through high-velocity capture, discover, or analyze process [3].