Industrial control systems (ICS) are used for remotely managing industrial installations and facilities. Supervisory control and data acquisition (SCADA) systems, distributed control system configurations, and other smaller control systems are all included in the term ICS [1]. These systems are widely used for large, distributed industries such as oil and gas, transportation, manufacturing, and electric power generation facilities [2]. As these mission-critical operations require constant monitoring, distributed control systems provide an advanced mechanism for the remote command and control of industrial plants and processes.

Distributed Control System refers to the division of a major application into smaller submodules where each module carries certain specific application processes and allows communication between these submodules. DCS is used for mission-critical industrial and manufacturing industries. It relies on decentralizing the control unit and establishing a shared network between the engineering stations. The DCS architecture consists of an engineering workstation, an operating station or Human Machine Interface (HMI), a process control unit or local control unit, intelligent devices, and a communication system make up. Sensors gather information and deliver it to input/output modules, where actuators help control process parameters [3]. The field control station receives input signals from the sensors. The system performs control calculations and outputs the signals to initiate compensatory actions. The field buses carry the results achieved upon processing and analyzing the input signals to the actuator devices. DCS distributes the control processing among the system’s nodes, thus resulting in a reliable, fault-tolerant system.

While the DCS are reliable, scalable, and fault-tolerant, there is a rising concern regarding the security and accuracy of the data transmission. Due to the architecture’s interconnectivity with the cyber world, it is prone to vulnerabilities, threats, and attacks. Some of the most common attacks on DCS-based environments include Distributed Denial of Service (DDOS), eavesdropping, man-in-middle, routing, malware, and Structured Query Language (SQL) injection attacks. The previous work discussed DCS technological environment followed by DCS environment-related threats and attacks [4]. This paper extends the last research work by further examining the risk management methodology by incorporating two additional components, namely, trust and reputation, where trust refers to the degree of confidence in the form of probability by which one entity will assume, or expect, the behavior of another entity [5–7], and reputation refers to a form of indirect trust where one entity relies on surrounding entities’ observation and opinion about the target entity’s behavior and reliability within a specified timeframe [8].

In this research we present a security evaluation scheme based on trust. The proposed TRUSED (Trust-based security evaluation scheme) intends to improve node trust by examining their direct and indirect interactions. The model distinguishes between untrustworthy nodes that should be segregated and trustworthy nodes that would be excellent for communication by combining direct and indirect trust.

In brief, our contributions can be summarized as follows:

This research discusses the most efficient classic and novel approaches for establishing and evaluating trust and reputation in the distributed environment.

The remainder of the paper is laid out as follows: The second section presents the most relevant previous work in the trust and reputation evaluation field. Section 3 covers the planned TRUSED scheme’s design, whereas Section 4 describes its implementation. Section 5 discusses simulation and experiment results, followed by Section 6’s concluding observations and recommendations for further work.

The trust and reputation management models can be categorized into two types: centralized architecture and distributed architecture. In a centralized architecture, one central entity is responsible for assessing trust across all the nodes within a network. This central entity acts as the trust manager for the network by maintaining all the trust scores and responding to all the incoming requests from the devices [13]. This approach is beneficial in reducing the overhead cost. Centralized architecture is not, however, very fault-tolerant. If the central entity fails to function, it will disrupt the entire network. In the distributed architecture, all the nodes are responsible for calculating the trust values and maintaining scores. Because of processing overhead associated with all the nodes and, unlike centralized architecture, the load is shared among all the nodes [14,15]. In addition to this, the failure of one node does not drastically affect the entire network, thus making it more fault-tolerant than a centralized network. The distributed control systems have a similar mechanism where the controllers are geographically scattered and each controller acts as a standalone communicator. Based on DCS functional similarities, this paper will focus on distributed reputation architecture for establishing trust between the service receiver and the service provider. This section presents the various computational trust and reputational-based models, recent advances in research effort, and their limitations.

Various novel and classic quantitative approaches exist for evaluating trust and reputation in a distributed environment. This section presents the various computational trust and reputational-based models, recent advances in research efforts, and their limitations.

The fuzzy theory-based trust and reputation model provides a mechanism to overcome trust establishment and management issues among CPS, devices, and wireless sensors, using reputation in the Internet of Things (IoT) [16,17]. The model focuses on sensor nodes that consider the quality of Service (QoS) metrics [13]. In this model, each node takes direct trust and indirect reputation to determine the trustworthiness of other sensor nodes [18]. Direct trust is derived from direct observations, and indirect reputation is developed based on the recommendations of neighboring nodes. Based on the nodes’ behavior in route discovery, maintenance, and data forwarding, the nodes are divided into two types, the malicious nodes that do not perform the package forwarding function and the nodes that do not participate in the route discovery phase.

The fuzzy model performs well in detecting the malicious nodes in Wireless Sensor Network (WSN). The model does not, however, address forged transactions and the effect of manipulated nodes. Thus, the model can be further optimized for discarding the influence of malicious nodes on the results and improvising the mechanism of updating trust.

On the other hand, Asiri et al. (2016) [17] propose a recommended-based trust and reputation model for overcoming the limitations of the Fuzzy theory-based trust model. The model protects against good-mouthing, bad-mouthing, and ballot stuffing attacks by constantly updating the weights and taking the history of behavior and rating quality into account for calculations. The model uses a probabilistic neural network framework to differentiate between trustworthy and malicious nodes [19]. The probabilistic neural network comprises a multi-layer architecture including input, hidden, pattern, and output layers [20].

The authors have introduced a distributed computing model where the computations are distributed between the nodes. The framework has defined the stronger nodes as the alpha nodes. These nodes are responsible for processing the computations [21]. The model also considers data sensitivity by taking an input parameter to specify the sensitivity level to differentiate between insensitive and confidential information and compute it accordingly [22,23]. It proposes a probability density function for calculating and distinguishing between untrustworthy and trustworthy nodes. Based on the probability evaluated by the neural function, the rating matrix is updated by the alpha nodes accordingly [22].

Overall, the model has guaranteed better availability, energy preserving mechanisms, reduced computation overheads over publicly available information [20], and protection against good-mouthing, bad-mouthing, and ballot attacks. The model, however, cannot protect against attacks such as distributed denial of service (DDOS), man-in-the-middle (MIM), and wormhole. Lastly, as per current research, the model lacks real-life implementation.

The quantitative model introduced by Yu et al. (2017) [24] emphasizes the importance of establishing trust relationships among sensor nodes by efficiently analyzing the nodes’ behavior. In this architecture, the main node calculates and analyzes the trust value to determine whether to transfer the packet to the next node or drop the packet [15]. The model considers packet forwarding capacity, repetition rate, consistency of the packet content, delay and integrity into consideration as trust factors [25,26]. The model utilizes the information entropy theory to determine each decision factor’s weight. The reason behind incorporating the information entropy theory is to eliminate the inconsistency of the decision model because of the subjective weight setting. In addition to this, the model has also incorporated Dempster-Shafer (D-S) theory to deduce and synthesize the trust for calculating the indirect trust values [20].

The quantitative model promotes secure packet forwarding and eliminates the inconsistency of the decision model because of subjective weight setting by using information entropy theory consumption. In addition, the model uses less energy due to trust exchange among neighboring nodes, thus reducing latency and energy. However, the research lacks practical needs-driven implementation. The research could be extended to integrate practical demands including efficient technologies and lightweight trust management for further improvising the framework.

The CTrust model analyzes the trust of a node based on its current and past interactions with a node [27]. The key features of this model include the introduction of practical solutions for concerns regarding the establishment of trust and reputation. The model comprises mathematical functions for trust assessment, delay recommendations, and aggregation for computing partial trust values [28,29]. According to the model, the trust is accessed based on current and past interactions and recommendations, although the weightage for the interactions differs. The current interactions weigh more than the past interactions. The model has defined the trust maturity threshold as 250 direct interactions, after which the two nodes can interact without a recommendation from other nodes. The trust parameters can be defined based on the context at the runtime, and the truster has the authority to assign the weight to each parameter [30–32].

This model has several benefits, including the weighted trust parameters, weighted recommendation function, trust decay function, and a method for developing trust maturity and equilibrium between two nodes. The trust decay function allows the past trust to degrade or be eliminated over time, thus allowing more weight to the new trust scores [33]. In addition, it addresses general self-promoting and opportunistic attacks [34]. However, the model lacks a threshold scale for the parameters. In addition, it requires more computing and energy resources, high energy consumption issue has remained unresolved. This concern has also been addressed by Wang et al. (2020) [30] and Azad et al. (2020) [35] in their comparative reviews. Lastly, the model has not addressed the data privacy concerns and the risks that could arise from security threats and attacks. Overall, the model has adapted innovative approaches for tackling trust and reputation development issues. However, there is one significant gap that the authors can address, which relies on parameters’ weight determination.

The Bayesian Model proposed by Teacy (2012) [9] places more weight on past interactions with trustees in a context like the present situation than on new interactions. The Bayesian model is a statistical trust and reputation model that relies on a beta probability density function based on interactions from neighbor nodes to obtain the preliminary information and direct interaction to get the posterior distribution to estimate the trust and reputation rating [36]. The Bayesian model is one of the most used architecture applications for the numerical aggregation of past interactions and statistical trust computation [9–12]. The model considers direct and indirect experience for evaluating the QoS parameters. The model comprises two components: the reputation model and the confidence model. The reputation model refers to the indirect experience where it considers the relationship between behaviors and observations of different models. The confidence model refers to the direct experience where the trustor’s perception about a trustee’s behavior is taken into consideration [37]. The benefits include the model’s ability to determine the expected behavior of a vaguely known agent by observing similar trustees and thus not restricting the agent to relying on mutually shared trustees’ observations. The feature helps the model overcome the agent’s whitewashing issue where the agent tries to incriminate facts to eliminate their bad reputation. Overall, the Bayesian model provides a well-defined framework that can be applied to a wide range of domain specific applications.

The predictive model defines an alternate approach for determining direct and indirect trust among nodes for routing packets between the service provider and receiver [38]. The model uses the beta probabilistic distribution approach for calculating the direct trust and AMRA/GARCH for determining the future behavior of the nodes based on their past performances. The authors have taken the number of packets properly forward: the number of packets dropped and the number of packets falsely injected are the main parameters for evaluating the model. In addition, a routing mechanism has been defined to ensure safe and accurate delivery of the packets by using the trustworthy nodes within the destination’s path [39]. The model is highly effective for predicting the trust values of multiple steps ahead in the series. The model efficiently detects the dropped or false injected packets and defines them as a blacklisted node. Most importantly, the model addresses various attacks including routing table overflow and resource consumption, DoS attack, sleep deprivation, and it also addresses spoofing attacks. However, this protocol alone is not suitable for highly dynamic networks. In addition, being a quantitative model, it cannot establish trust unless some behavior response gets recorded. Therefore the model requires vulnerable time to detect malicious activities.

Every trust and reputation model targets a set of risks and attacks associated with the distributed environment. When selecting the aggregation technique for the trust model, it is important to determine the major attacks that the proposed trust model must address to assist in mitigating the target risks associated with the environment. Table 1 summarizes the comparison of various trust and reputation models.

A trust-based security evaluation scheme (TRUSED) has been proposed and presented in this section. The components of TRUSED are discussed in the subsequent sections, where the following assumptions are made:

Nodes are deployed randomly, without mobility, and communicate via a shared bi-directional wireless channel within their communication range.

The efficacy of the proposed TRUSED scheme is assessed using the OMNET++ simulator due to its open-source and discrete nature with rich support of the graphical representation of a network. The nodes are deployed randomly within the network area of 100 m × 100  m. Furthermore, the simulation time for different studies range from 200 to 1200 s. A constant bitrate (CBR) traffic generator with a packet size of 50 bytes is utilized with the transport layer protocol (UDP). TRUSED’s results are analyzed for trustworthiness, malicious node detection rate, detection accuracy, packet drop ratio, average network throughput, and packet delivery ratio.

IEEE 802.15.4 was utilized as the physical layer standard, and UDP was used as the transport layer protocol, with a constant bit rate (CBR) traffic generator and a data packet size of 50 bytes. The simulation length ranges from 200 to 1200 s, with a variable number of network nodes ranging from 10 to 50, randomly distributed over a 100 m × 100  m area. Table 2 shows a summary of simulation parameters.

Moreover, Fig. 3 depicts the flow chart for the proposed TRUSED.

The algorithm for node trust calculation is as follows:

The algorithm steps are explained as follows:

Step 1: The algorithm takes the nodes’ behavioral characteristics such as received packet assessment, send packet assessment, and transit packet assessment as input values.

To evaluate the performance of the proposed TRUSED scheme various performance parameters were considered and benchmarked against GDTMS [32] and CTrust [27], which are as follows:

Impact of trustworthiness: This parameter depicts the impact of the trustworthiness level over time and in the presence of malicious nodes in the network.

Securing the Distributed Control Systems (DCS) is challenging due to their pervasive environment and complexity. Traditional network security mechanisms are ineffective regarding comprehensive solutions in securing DCS since they cannot detect the malevolent nodes. Trust-based security measures are helpful and substantial solutions and play a vital role in securing the DCS environment. In this paper, we proposed the TRUSED scheme that differentiates between the trustworthy and the untrustworthy nodes for promoting secure communication within a distributed environment. This scheme is designed to enhance security for Distributed Control Systems by monitoring the nodes and isolating the malicious ones. TRUSED is evaluated against a various performance characteristics, including trustworthiness, malicious node detection rate, detection accuracy, packet loss ratio, packet delivery ratio, and average network throughput, using the OMNET++ simulator. The suggested scheme’s outcomes are compared to two existing models, GDTMS [32] and CTrust [27]. Overall, TRUSED performed better when compared to the benchmark criteria.