This research leverages the contributions of a few relevant works through a thorough review of their methods, implementations, and limitations. To ensure a comprehensive review, specific keywords and logical operators were used in searching through academic databases and search engines like “Science Direct,” “Mendeley/Elsevier,” and “Google Scholar” respectively. These keywords include the following: “Internet of Things OR Web of Things,” “Robotics AND its applications,” “Covid-19 AND Robotics”, “Application of Robotics in Post-Covid-19”, “Internet of Robotic Things,” “Internet of Things OR Internet of Robotic things AND Security,” “Network Security for Internet of Things,” “Artificial Intelligence OR Machine Learning AND Internet of Things.” This approach offered a lot of insight into all the fundamental aspects of the proposed work, especially towards identifying the most efficient methodology for realizing the research goals. Using a year filter also ensured retrieving only the most relevant results since 2020.
The works of [2][11] introduce the fundamental concepts and definitions of the Internet of Things (IoT) and the Internet of Robotic Things (IoRT), respectively. In [2], IoT is defined as the interconnection of various sensor-based devices as part of a networked ecosystem to enable information and resource sharing for enhanced service delivery. In their view, IoT is already influencing the world through various current and proposed applications. For example, a smart fridge (IoT example) can have communication access to a cloud server for purchasing depleted groceries through available e-commerce platforms. In the context of this example, the smart fridge may not necessarily have the computing power to navigate e-commerce websites to make the needed purchase; however, it can communicate small information or a list of needed items, its unique ID, address, etc., which is then processed via the cloud server.
While IoT is fascinating and projected to save more time and resources, the most prevalent challenge for this new technology is the security of devices and information sharing on any IoT dedicated network. Typically, security is a major consideration for any new technology meant for commercial or global adoption. However, the concerns about the security of IoT devices become even more pertinent since a majority of these devices will interact directly with people in mostly private or corporate settings [12][13]. Most sensor-based devices may have either video or audio recording features that make them vulnerable as listening devices for malicious actors. Also, devices that leverage the power of cloud computing usually connect and interact with a central server, which risks compromising the entire network of devices in a successful server hack. These security considerations and similar issues must be resolved before adopting IoT across a wider range of devices for personal or enterprise automation. Other challenges mentioned by the authors include “The cost of IoT implementation” and “connectivity issues for areas” with poor network coverage.
IoT provides a conceptual infrastructure for many other applications across various industries like IoRT (Internet of Robotic Things), IoMT (Internet of Medical Things), V2X (Vehicle to Things Communication), and Internet of Battlefield Things (IoBT), etc. However, the variations in this technology are emphasized at the Application Layer of the framework. This is described in [11] that the Internet of Robotic things is just an IoT network where the networked devices are robots and other similarly autonomous systems that facilitates the automation of complex processes and resource sharing. Typically, this network accesses a central repository (such as a cloud or local server) for leveraging advanced computational power, big-data storage of activity logs and resources like software and firmware upgrades/updates, and remote control of multiple robotic devices using a group policy. The authors discussed the distinction between the Internet of things (IoT) and the Web of Things (WoT) as the overall framework's network and application layers. In this context, IoT provides a network platform for devices from different manufacturers to enhance their functionality by fetching and using data available via a connected network (typically the internet). In Web of Things (WoT), the connected devices are typically a similar architecture, from the same manufacturer, or share the same control interface. Depending on the intended level of network integration or exposure, IoT or WoT may be appropriate for synchronizing a network of robots and robotics devices [11]. The feasible approach for an IoT implementation is enabling a secure API for authenticated communication between cross-platform devices or “things” [14]. This reduces the risk of a coordinated attack that preys on insecure networked devices to infect others [6]. WoT may only need to integrate an authentication framework in the communication channel due to the expected relative homogeneity of the networked devices. Irrespective of implementing these concepts as IoRT or WoRT, this security challenge, among others, is still prevalent. Also, this challenge is peculiar to all the applications of the IoT framework, such as IoRT, discussed in this paper. However, IoRT is being deployed with consideration for security as part of the “Industry 4.0” framework.
In [5], “Industry 4.0” is introduced as the new industrial revolution, first coined in Germany to describe the future of industrial automation, production, and efficiency. Here, the authors present an overview of the applications of IoT technology in realizing the ideal Industrial revolution through truly automated processes via a network of autonomous or robotic devices. This ensures that all major aspects of industrial activities run at all times, irrespective of the presence of human labor. In this framework, the human workers perform the role of supervisors to ensure that the executed or progressing tasks handled by the robotic devices are as expected.
The works of [15] and [4] also define the architecture of the Internet of Robotic Things as composed of five (5) separate layers: the hardware/physical, network, internet, infrastructure, and application layers [15]. Each layer has its security vulnerability; however, the network/internet layers are most prone to attacks via access points or compromised IoRT devices. For IoRT, in the context of specialized industrial deployments or applications, the other aspects of the technology can be protected by using trusted and customized hardware, infrastructure, and secure application development methodologies. However, the obvious solution for safeguarding the network/internet layer is to design and implement a smart and well-tuned firewall to filter content uploaded and downloaded from the local network or internet in the case of a global IoRT scheme [16]. Since the IoRT is a generalization of IoT, the same (or slightly modified) measures for security applications to ensure the integrity of communication between device endpoints within the network. Hence, this work considers security approaches available through IoT and IoRT research.
One such security measure is reviewed in [17], where Blockchain technology is proposed as a solution for ensuring security and transparency of data transmission within an Internet of Industrial things (IIoT) network. The authors also used “Industry 4.0” as a case study where industrial systems are interconnected across manufacturers and suppliers for smoother integration of the industrial process. They identify “trust between connected devices” as the key to ensuring safe information exchange and transactions within the proposed industrial network. Their proposed Blockchain technology with Smart Contracts is based on the discussions presented in [18] on the applications of Blockchain technology, especially the platform that powers the popular decentralized cryptocurrency “Bitcoin.” Smart contracts are immutable self-enforcing agreements/contracts implemented in code on a Blockchain, i.e., Ethereum, triggered by a set of defined events or actions [19]. They are essentially digital versions of traditional contracts without a need for an authenticating third party or enforcer, i.e., a court, law enforcement, banks, etc., as in traditional contracts or agreements. In the context of their application, industrial entities or machinery within a designated production process can share information about process status instantly via smart contract technology. In this scenario, a transaction occurs when an industrial device signals its completion (thereby initiating a sub-process completion transaction). The next associated industrial device acknowledges the initiated transaction to begin its process. This exchange is useful for the speed of autonomous actions of smart industrial devices. It leverages the security of Blockchain technology to ensure that transactions are immutable and only invoked on the dedicated private or public Blockchain platform. Also, the decentralization of the Blockchain ensures that instructions are adopted on a consensus basis. This further limits the activities of hackers or other malicious third-party actors from tampering with information passed to the connected devices. While this is a positive outcome, a significant amount of computation is required per network node to keep a Blockchain application sustainable. Hence, implementing this approach will lead to an expensive and energy-inefficient system.
Given the inherent computational cost and energy inefficiency of implementing a Blockchain-powered IoRT network, other modern technologies are presented through the works of [20], where the authors compared the efficiency of Blockchain, Fog computing, Edge computing, and Machine Learning for IoT security. Regarding functionality, Edge and Fog computing are essentially the same technology. Implementing an Edge or Fog computing methodology aims to reduce processing and response latency for devices requiring high-speed real-time communication with a central server or data center [21][3]. These network architectures reduce the reliance of the networked devices on a cloud server by enabling the performance of computations via a local edge or fog server within the network. Pertinent computations that may exert undue stress on the hardware of the local server are then pushed to the associated cloud service. This reduces the traffic of requests hitting the cloud service for mundane commands or instructions, and it also limits the likelihood of a DDoS attack which can be deployed through the network of devices or the cloud server [22]. Unlike Blockchain technology, deploying a Fog or Edge computing architecture ensures the nature of a typical IoT or IoRT network is preserved with minimal to no computation done by the client devices [23]. However, significant redundancy of Fog/Edge computing nodes may be required to ensure multiple points of failure in the networking scheme. Depending on the architecture implemented, this may be costlier than paying for cloud resources. The security, upload, and backups of each fog or edge server are also the responsibility of the network or system administrator managing the local edge or fog servers. The operational costs are also higher since a certain level of experience is required to administer these complex networks. Hence, a cloud computing scheme in which the selected hosting data center is within proximity of the organization or network may be a preferable option. In the case of robotic things, device-based edge computing can be implemented to leverage the already available computing hardware of robotic devices, and the cloud server is only accessed when necessary. Irrespective of the scheme of choice, an adaptive security scheme is recommended to handle known or newer threats to the network and its associated devices. In that regard, machine learning techniques and algorithms are explored for identifying common and new network threats.
Machine learning is an aspect of Artificial Intelligence in which computer systems are trained to perform cognitive tasks like pattern recognition, image classification, etc., using knowledge gained from iteratively mapping an input set of problem-specific data to a corresponding output model for applications in intelligent decision making and adaptive predictive or forensic analytics. In essence, Machine Learning allows computer systems to find solutions to problems without explicit programming [24][2]. Similarly, Machine Learning can be implemented in a more immersive form to achieve “Deep Learning.” Deep Learning models are Artificial Neural Networks with many hidden layers and advanced activation functions for extensive back-propagation and error correction. These types of algorithms intuitively require a lot of computing resources. They are typically applied to solving problems that require a very high level of accuracy and use high-dimensional data [25]. Hence, there is usually a trade-off between implementing standard Machine Learning algorithms and Deep Learning algorithms in terms of computational resources, model accuracy, and training data volume and dimensions [24][3][4]. In the context of threat identification for authentication of commands and data interaction within an IoRT network, Machine Learning techniques for pattern recognition or classification like Support Vector Machines, Decision Trees and Random Forest Classifiers, Artificial and Deep Neural Networks, etc. are most suitable [26][27].
Essentially, the model serves as a “network administrator” that iteratively monitors the network traffic within a deployed Fog, Edge, or Cloud-based IoRT architecture to ensure a malicious attacker does not compromise the enrolled robotic systems. This approach is described in the works of [28], where a machine learning-based security system, powered by a Recurrent Neural Network (RNN) with a Long-Short-Term-Memory (LSTM) hidden layers, is developed for detecting malware in a network of Android OS-based IoT devices. According to the authors, the Android OS is not always secure against malware and can lead to vulnerabilities in an IoT network. In their approach, the trained RNN-LSTM-based-network-analyser scans device traffic on the network for malware patterns, and a recommended action is passed to the network firewall to either allow or block that traffic. The N-Adam optimization algorithm was used in their implementation to train the developed Neural Network on the most relevant networking features available in the training dataset. This allowed for faster computation and adaptation of the learning algorithm when new malware-definitions data needed to be trained. Since every machine learning task requires a dataset for training and testing, the authors used two separate malware-benign IoT-traffic datasets of varying volumes described in [29] and [30] (IoT-23 dataset) to test their proposed algorithms in comparison with commonly employed machine learning models. According to their results, the proposed RNN-LSTM model provided excellent results compared to other implemented models. However, the model results in increased computational requirements if deployed on individual IoT devices, as described by the authors.
The IoT-23 dataset was also used by [31] in training and testing various machine learning algorithms for detecting attacks in an IoT network. Support Vector Machines (SVM), Naïve Baiyes (NV), Random Forests (RF), Decision Trees (DF), and Logistic Regression (LR) were implemented to classify the network traffic into one of twelve (12) categories/labels (one of the categories being an attack signal). The classifiers achieved a decent level of threat identification, with the Decision Trees and Random Forest classifiers achieving the highest rates.
Another mitigation strategy is presented by [6] for network security in the presence of vulnerable devices. This strategy is termed the IoT sentinel, and it identifies devices on a typical IoT network while enforcing security policies on vulnerable devices. The developed system controls traffic flow from vulnerable devices to ensure they do not allow a breach of the network. The device identification is achieved via a Random Forest machine learning algorithm, and the network security rules are enforced by a Security Gateway through which every device authenticates on the network.
A survey of various machine learning and deep learning methods is given in the works of [10], where several approaches for identifying malware are considered. Machine learning methods like Decision Trees, Support Vectors, Naïve Bayes, K-Nearest Neighbours, etc., are presented along with Deep Learning methods such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). Unsupervised approaches like K-Means, Genetic Algorithms, and Principal Component Analysis (PCA) are also introduced for IoT Security applications. Due to each algorithm's expected computational complexity, an edge-connected network running such an algorithm is considered sub-optimal. Also, GPU processing is suggested for speeding up the training and deployment of learned models for IoT security. However, this is generally achievable via a cloud-edge-based implementation that reduces computational complexity on the edges. The review did not conclude on the relative advantages of using a deep learning approach over a conventional machine learning technique; however, the works of [32] inform that deep learning provides better detection results for IoT security applications than conventional machine learning algorithms. Also, they propose an offloading strategy for optimizing the computational complexity of deep learning networks with edge computing. These and many more approaches proposed through the reviewed literature inform various algorithms and their relevance to the task. The major limitation identified is the extra computational overhead required for implementing most of the discussed algorithms. However, the minimal resources available to control systems of a typical set of robotic things constrain computational efficiency on the optimal approach for securing an IoRT network. This is factored into our proposed methodology and choice of machine learning techniques.