A Deep Reinforcement Learning–Based Trust Management Model for Secure Routing in Vehicular Ad-Hoc Networks under Real-Time Traffic Variations

Smart transportation Vehicular Ad-Hoc Networks (VANETs) are the state-of-the-art technologies, which offer on-board functions of decentralized wireless communication with vehicles and roadside equipment. Traffic safety alerts, congestion avoidance, cooperative driving, and the proliferation of infotainment are among the most essential services provided by the networks. However, VANETs also possess certain operational and security-related drawbacks that do not allow them to be trusted in the real world. The topology of the network is dynamic, and the high level of mobility of the vehicles usually disrupts the path of routes, leading to unstable connections, packet loss, and poor quality of service (QoS). The existing routing processes are not able to make quick changes in topology in direct relation to the timely delivery of information and the resiliency of the network as a whole. Other than the mobility issues, VANETs are also very susceptible to security threats. In practice, rogue or malicious nodes may be introduced into the network and cause counterfeit routing, selectively drop packets, or launch a denial-of-service attack. However, these actions not only impact the routing reliability but also compromise trust between participating nodes, which has an impact on network performance and causes the network to become even more vulnerable. All these challenges lead to a major problem, which is that current VANET routing options are unable to react dynamically to the state of the network and avoid trust-based and routing-level attacks [1, 2].

Conventional routing algorithms such as Ad hoc On-Demand Distance Vector (AODV) routing protocol, Dynamic Source Routing (DSR), and Optimized Link State Routing (OLSR) protocol are well-suited to the dynamics of mobile ad hoc networks, but do not perform well in VANETs, as the rates of route breakage and message overhead are too high. Such approaches overlook critical factors of trust and fail to learn in real time from traffic behavior, thereby subjecting them to security and performance bottlenecks [3, 4]. This has led to new trends in VANET research interest, shifting toward the use of trust routing systems in which nodes store and maintain trust scores for direct and indirect encounters. The functionality of these systems is to identify and quarantine malicious nodes, thereby strengthening secure transmission paths within the network. These trust systems are deemed reactive and, more often than not, inherently static in their formation, offering limited flexibility for mobility in high-speed, highly variable environments [5].

To support the demand for intelligent, adaptive, and trust-aware routing in VANETs, Deep Reinforcement Learning (DRL) has been recognized as a valuable paradigm. DRL agents can learn the best policy through repeated interactions with the environment, and the feedback is in the form of a reward or a penalty that guides their decisions. In Contrast to typical machine learning models, which can operate only on predetermined datasets, DRL operates in an online learning environment and is therefore applicable to dynamic systems, such as VANETs [6]. Compared to other DRL architectures, both Asynchronous Advantage Actor-Critic (A3C) and Dueling Deep Q-Network (D-DQN) are promising methods for large-scale decision-making and high-dimensional tasks. Such architectures either separate value and advantage estimation or actor and critic learning, respectively, which reduces the rate of convergence and improves the robustness of their policies [7, 8].

Regarding VANET routing, DRL agents can discover the best routes not only by focusing on standard metrics such as latency and hop count but also on security metrics such as trust scores. Trust scores are then calculated dynamically depending on a compilation of the past activities, packet forwarding integrity, and motion steadiness. The nodes that are suspicious in nature, like packet dropping or altering the content of messages, lose their trust value over time, which directly affects routing decisions. This way, the integration of DRA and trust management has resulted in a smart, flexible routing structure that is secure and performance optimized [9].

Although this approach has great potential, implementing DRL-based trust routing models in real-time VANET environments poses several practical challenges. Most of the existing implementations are limited to simulated or synthetic configurations and non-real-time responses. In addition, in most models, trust behaviors are treated as static or semi-static, failing to account for rapidly evolving attack patterns or context-sensitive routing choices. Moreover, the comparative behavior of the DRL models in this context, considering both real traffic patterns and adversarial threats, has not been deeply explored [10].

To address these challenges, the study introduces a DRL-based routing protocol agent, closely coupled with a real-time trust management technique. The node behavior profiles are continuously used to update the trust scores, and by leveraging these trust scores and the network parameters, the DRL agents select the most reliable routes. Two individual DRL architectures, namely D-DQN and A3C, are applied and compared, and tested across different vehicular settings. Performance is determined using the metrics of packet delivery ratio (PDR), average latency, energy efficiency, and stability of trust convergence that can be used to estimate the applicability of each model to different network loads and threats.

Figure 1 is a graphical summary of the proposed architecture. It displays vehicles moving on a highway, and the trust score of each vehicle updates in real time based on its behavior. Malicious nodes are red nodes, and they can drop packets and misroute. It uses each vehicle's trust parameters and real-time state features (e.g., delay, route stability) to compute the best path, assisted by a central DNR decision engine (depicted as a controller). The green lines represent safe lines that have been gained through DRL policies, and the red lines are unsettled or malevolent paths.

Figure 1. Deep Reinforcement Learning (DRL)-integrated trust-aware secure routing architecture for real-time Vehicular Ad-Hoc Network (VANET) environments

The framework has four crucial components, namely, creation of a trust score, working with the DRL model, and adversarial detection. It describes how security and route optimization are integrated in real time. The proposed solution is a step towards enhancing the state-of-the-art through integrating trust computation and DRL in one secure routing framework. In contrast to traditional trust systems, which rely on periodic evaluation, the proposed model adopts a continuous learning approach using DRL agents and is therefore real-time adjustable. Additionally, comparing D-DQN with two routing protocol versions (DSR and A3C) through benchmarking provides this work with a significant basis for choosing the right models, depending on the VANET deployment scenario and resource limitations.

The nature of VANETs is highly dynamic and security-conscious, as traffic conditions frequently change with the evolving topology, connectivity is intermittent, and malicious nodes may exist. The current routing protocols, most of which are complemented with trust-based mechanisms or reinforcement learning (RL) solutions, are not sufficient to ensure a secure, latency-free, and efficient communication path. Most of the strategies are passive and inactive in managing trust or handling dynamic problems arising from increasingly real-time traffic. Moreover, previous studies did not conduct a systematic analysis of the comparative resilience and performance of advanced DRL models for trust-aware VANET routing, such as D-DQN and A3C. Therefore, there is an important requirement for a smart, trust-encompassing routing system that can update routing in response to real-time traffic and threat scenarios while maintaining a good QoS. In short, secure routing in VANETs must satisfy mechanisms that are not only intelligent and adaptable but also aware of security threats and able to operate within time constraints. These requirements are fulfilled through DRL, which combines active learning and behavior-based trust measurement into an overall routing strategy. The given architecture is tested under real-time traffic conditions to demonstrate its performance and efficiency in defending vehicular communication. This work has two principal contributions as follows:

• A secure routing scheme based on DRL is designed, specifically, the D-DQN model, which aims to optimize routing adaptively for an adversarial environment.
• A trust scoring mechanism is implemented into the DRL agent to enhance the reliability of nodes depending on the analysis of their behavior patterns and latency.
• To measure performance, a comparative analysis is made against DSR and A3C models within the same simulation environment.
• The natural modeling of traffic flow and communication in an urban mobility scenario is achieved by the use of a co-simulation environment that integrates Simulation of Urban Mobility (SUMO) and Network Simulator (NS3).

The rest of the paper is divided into the following sections: In Section 2, related work is presented, describing conventional and AI-based routing and trust mechanisms in VANETs. In Section 3, the problem statement and research objectives are illustrated. Section 4 describes the proposed methodology architecture, DRL formulation, and reward functions. Section 5 gives a detailed experimental explanation, indicating simulation configurations and modules. The analysis and comparison, presented using figures and tables, are given in Section 6. Lastly, Section 7 concludes the paper and outlines what can be done next.

The scheme in the proposed intelligent motor vehicle routing architecture for VANETs uses vehicle behavior to assess trust and DRL to achieve real-time optimality, security, and adaptability of path routes. Trust Score Calculation, DRL Policy Learning (D-DQN, A3C), and Secure Route Execution are the most central parts of the architecture. Synergistically, these components are designed to ensure secure and efficient data transmission, despite the inevitable presence of malicious nodes.

3.1 Trust evaluation mechanism

Each node computes a trust score ${{T}_{i}}$ for its neighbors based on three primary behavior metrics: Packet Forwarding Ratio (PFR), Communication Consistency (CM), and Mobility Honesty (MH). These indicators are continuously monitored during communication and mobility as illustrated in Eq. (1) [24-26].

${{T}_{i}}={{w}_{1}}\cdot PF{{R}_{i}}+{{w}_{2}}\cdot C{{M}_{i}}+{{w}_{3}}\cdot M{{H}_{i}}$        (1)

where,

• ${{w}_{1}},{{w}_{2}},{{w}_{3}}\in \left[ 0,1 \right]$ are normalized weights such that ${{w}_{1}}+{{w}_{2}}+{{w}_{3}}=1$.
• $PF{{R}_{i}}=\frac{Packets\_Forwarde{{d}_{i}}}{Packets\_Receive{{d}_{i}}}$
• $C{{M}_{i}}$ evaluates signal stability over time.
• $M{{H}_{i}}$ penalizes erratic GPS movement behavior.

To ensure that trust scores reflect time-evolving behaviors, they are dynamically updated using an exponential decay function as shown in Eq. (2):

$T_{i}^{(t+1)}=\delta \cdot T_{i}^{(t)}+(1-\delta )\cdot {{T}_{\text{new}}}$     (2)

where, $\delta $ is a memory factor controlling how past behavior influences the new score.

3.2 Reward function for Deep Reinforcement Learning agents

The DRL agents learn to select optimal routing paths based on a reward signal that encapsulates security and performance objectives [27]. The reward ${{R}_{t}}$ at time $t$ is defined in Eq. (3):

${{R}_{t}}=\alpha \cdot \text{PD}{{\text{R}}_{t}}-\beta \cdot {{L}_{t}}-\gamma \cdot {{E}_{t}}$       (3)

where,

• $\text{PD}{{\text{R}}_{t}}$: Packet delivery ratio.
• ${{L}_{t}}$: End-to-end delay.
• ${{E}_{t}}$: Energy consumed per transmission.
• $\alpha ,\beta ,\gamma $: Reward weights selected based on environmental goals.

This equation reinforces paths with higher delivery success, lower latency, and minimal energy usage.

3.3 Dueling Deep Q-Networks Q-value computation

In Dueling DQN, the action-value function is decomposed into two streams of state-values, and the advantage is to stabilize training and reduce overestimation as defined in Eq. (4) [22, 23]:

$Q\left( s,a \right)=V\left( s \right)+A(s,a)\frac{1}{\mid A\mid }\sum\limits_{{{a}^{\prime }}}{A(s,{{a}^{\prime }})}$        (4)

where,

• $V\left( s \right)$: Value function representing the quality of the state.
• $A\left( s,a \right)$: An advantage function estimates the relative utility of an action $a$.

This approach enhances decision-making by focusing on advantageous actions under uncertain or similar state conditions.

3.4 Asynchronous Advantage Actor-Critic policy gradient formulation

In the A3C model, multiple asynchronous agents run in parallel to update a shared policy network. The actor’s policy is optimized using the gradient shown in Eq. (5):

${{\nabla }_{\theta }}J\left( \theta  \right)=\mathbb{E}\left[ {{\nabla }_{\theta }}\text{log}\pi ({{a}_{t}}\mid {{s}_{t}};\theta ){{A}_{t}} \right]$       (5)

The advantage function ${{A}_{t}}$ is estimated as Eq. (6):

${{A}_{t}}={{R}_{t}}-V\left( {{s}_{t}} \right)$          (6)

where, $V\left( {{s}_{t}} \right)$ is the critic-estimated value for the current state. This method improves training stability and responsiveness to dynamic conditions.

3.5 Latency modelling

The average latency ${{L}_{avg}}$ for a path is computed as defined in Eq. (7):

${{L}_{avg}}=\frac{1}{N}i=1NLi+Qi$       (7)

where,

• ${{L}_{i}}$: Transmission delay at the hop $i$.
• ${{Q}_{i}}$: Queuing delay at the hop $i$.
• $N$: Total number of hops.

This allows the routing algorithm to minimize total transmission delay during path selection.

3.6 Energy consumption equation

Energy efficiency is a vital concern for VANET routing. The total energy consumed per route is expressed in Eq. (8):

${{E}_{p}}=\underset{i=1}{\overset{N}{\mathop \sum }}\,\left( {{E}_{tx}}(i)+{{E}_{rx}}(i) \right)$         (8)

where, ${{E}_{tx}}\left( i \right)$ and ${{E}_{rx}}\left( i \right)$ represent the energy to transmit and receive at the node $i$.

3.7 Deep Reinforcement Learning training loss function

For the DQN family, the loss function $L\left( \theta  \right)$ used to update network weights is the Mean Squared Error (MSE) between predicted and target Q-values, as defined in Eq. (9):

$L\left( \theta  \right)=\mathbb{E}\left[ {{(r+\gamma \underset{{{a}'}}{\mathop{\text{max}}}\,Q({s}',{a}';{{\theta }^{-}})-Q(s,a;\theta ))}^{2}} \right]$         (9)

where, ${{\theta }^{-}}$ refers to the target network parameters, updated at fixed intervals to stabilize training.

3.8 Network stability index

Route reliability is evaluated using a stability index ${{S}_{r}}$ Eq. (10):

${{S}_{r}}=\frac{\text{Link}\ \text{Up}\ \text{Time}}{\text{Simulation}\ \text{Time}}$     (10)

Higher values of ${{S}_{r}}$ indicate more durable links suitable for sustained communication.

3.9 Trust-weighted next hop selection

When making routing decisions, trust scores are used to filter and prioritize potential neighbors, Eq. (11):

$nh=\text{arg }\underset{j\in \mathcal{N}}{\mathop{\text{max}}}\,\left( {{T}_{j}}\cdot Q(s,{{a}_{j}}) \right)$         (11)

where, $\mathcal{N}$ is the set of neighboring nodes. Nodes with low trust are naturally avoided, even if they provide shorter paths.

3.10 System architecture diagram

The architecture (Figure 2) of the proposed secure VANET routing framework is depicted below using DOT code. It outlines the interaction among the trust management layer, the DRL agent, and the routing decision module in a vehicle node.

All the vehicle nodes have a Trust Evaluation Module that constantly evaluates their neighbours. The DRL Agent uses the trust scores and features of the environment (e.g., delay, hop count) to make an informed choice. The Routing Execution Unit then uses the selected next hop to safely send packets.

Figure 2. Layered architecture for Deep Reinforcement Learning (DRL)–integrated secure routing in Vehicular Ad-Hoc Network (VANET)

3.11 Deep Reinforcement Learning–trust routing algorithm

VANETs' architecture requires vehicle-to-everything (V2X) communications, which entails trust-aware routing because vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) connections evolve quickly with mobility and traffic density, and/or changes in wireless channel state, and fixed routing trees or fixed metrics are no longer trusted. Meanwhile, the network also becomes susceptible to misbehavior or malicious nodes, which can drop, slow down, or alter packets. The DRA agent can continuously update optimal next-hop solutions in dynamic environments by maximizing multiple objectives, such as delivery success, delay, energy consumption, or channel utilization. Nevertheless, without modelling trust, even DRL can select high-performance links that traverse untrusted nodes. The addition of trust provides an explicit security/reliability level that grades nodes by the observed behaviour during forwarding and consistency to enable the routing policy to favour nodes that are efficient and trustworthy, and enhances resilience to attacks (e.g., blackhole/grayhole, selective forwarding, and delay manipulation) and route failures due to unreliable participants. The following algorithm implements the DRL-trust-aware routing.

This trust-based DRL routing algorithm is trained to select a safe next hop that balances reliability and efficiency across a VANET graph, node strategies, and traffic. At every simulation time step, the agent constructs a state $a_{\mathrm{t}}$. Once it has sent the packet, it observes the behavior of the chosen node (successful forwarding or drop, additional delay, and energy cost) to generate trust evidence $E_t\left(a_t\right)$, which is used exclusively to create trust evidence. $\mathrm{T}_{\mathrm{at}} \leftarrow \delta \mathrm{T}_{\mathrm{at}}+(1-\delta) \mathrm{E}_{\mathrm{t}}\left(\mathrm{a}_{\mathrm{t}}\right)$. Then, a reward rt is calculated to accelerate delivery and trusted forwarding, and to deter delays and energy use. The DRL model is updated with the transition $\left(\mathrm{s}_{\mathrm{t}}, \mathrm{a}_{\mathrm{t}}, \mathrm{r}_{\mathrm{t}}, \mathrm{s}_{\mathrm{t}+1}\right)$. Repeatedly doing this will condition the agent to shun bad or unreliable nodes, resulting in a secure, optimal path to the destination.

3.12 Flowchart of operation

The flowchart in Figure 3 is a high-level description of how the secure routing process will occur at every node. It starts with the initiation. At each time step, the system scans the surroundings, selects a path, sends data, and boosts trust and enhances the learning model. This cycle allows the system to remain dynamic and adjust to the evolving conditions of the vehicles.

Figure 3. Deep Reinforcement Learning (DRL)–based trust-aware secure routing workflow in Vehicular Ad-Hoc Network (VANET) simulation

3.13 Dataset, simulation and evaluation details

1. Dataset: Mobility traces are generated using SUMO over a 2000 × 2000 m² urban grid, simulating realistic vehicle movements with speeds between 30–60 km/h.
2. Communication Protocol: NS-3's IEEE 802.11p MAC + DSR/DRL routing layer.
3. Simulation Time: 200 seconds per run, 10 independent seeds.
4. Evaluation Metrics:

1. PDR
2. Latency (end-to-end)
3. Energy Consumption (mWh)
4. Trust Score (stability and convergence)
5. Policy Convergence Speed

5. Implementation:

1. DRL Models implemented in PyTorch
2. Integrated with NS-3 using OpenAI Gym interface
3. DRL hyperparameters:

• Learning rate: 0.0005
• Discount factor ($\gamma$): 0.95
• Replay buffer: 10,000
• Batch size: 64
• Exploration: $\varepsilon$-greedy (D-DQN), entropy regularization (A3C)

The effectiveness of the suggested DRL-based, trust-sensitive secure routing scheme with VANETs was tested in dynamic adversarial simulation scenarios. An in-depth performance analysis was done to compare D-DQN, A3C, and the traditional DSR routing protocols. Each of the models was subjected to the same pattern of vehicular mobility, node densities, and threat conditions.

The metrics important during the evaluation included PDR, Latency, Energy Consumption, Routing Overhead, Trust Convergence, Path Stability, Reward Accumulation, and Malicious Node Isolation Accuracy. The simulation involved 120 vehicular nodes operating for 500 seconds, with 20% of the nodes behaving maliciously. Observations were logged every 10 seconds.

5.1 Packet delivery ratio

Figure 5 shows that D-DQN outperforms A3C and DSR in terms of PDR. The final simulation display showed an average delivery rate of 93.2 percent for DQN, 89.4 percent for A3C, and 81.6 percent for DSR.

The advantage-value decomposition provided by D-DQN improved its performance in changing trust (especially with swift topology fluctuations) (as shown in Table 3).

Figure 5. Packet delivery ratio (PDR) over simulation time

Table 3. Packet delivery ratio (PDR) over time intervals

Note: D-DQN = Dueling Deep Q-Network; A3C = Asynchronous Advantage Actor-Critic; DSR = Dynamic Source Routing.

5.2 Latency analysis

Figure 6 represents end-to-end latency. Although A3C initially achieved higher latency, its performance deteriorated under a heavier network load. D-DQN recorded an average latency of 110 ms, while those of A3C and DSR were 117 ms and 138 ms, respectively, with many route discoveries.

Figure 6. End-to-end latency comparison

5.3 Energy consumption

D-DQN consumed less energy (avg. 364 mWh) than other algorithms, which required fewer retransmissions and smart trust-based routing, as shown in Figure 7. A3C achieved a saving of 337 mWh, whereas DSR consumed the most, 402 mWh, due to route instability and re-broadcasts.

Figure 7. Energy consumption per packet

5.4 Trust convergence

The issue of trust convergence (Figure 8) shows that D-DQN effectively isolates bad nodes in a short time. D-DQN trust scores became stable by episode 50, A3C convergence stabilized by episode 80, and DSR had no trust mechanism in place, making it (the overall system) susceptible to attacks over time.

Figure 8. Trust score evolution for benign and malicious nodes

5.5 Routing overhead

In D-DQN, the normalized routing overhead (Figure 9) was the lowest because it achieved effective reuse of letters of trust. A3C incurred moderate overhead, whereas DSR incurred the largest overhead because many control messages were exchanged during route breaks.

Figure 9. Routing overhead vs. time

5.6 Path stability

The D-DQN demonstrated high path stability and a lower number of route breaks, as illustrated in Figure 10. There was an intermediate path churn with A3C. DSR experienced the problem of expensive route rediscoveries, which had a serious impact on data continuity.

Figure 10. Path stability over simulation time

5.7 Malicious node isolation accuracy

D-DQN achieved the best black hole isolation accuracy of 96.3, and A3C came next at 91.4. The misrouting persisted, and DSR had no isolation mechanism (Figure 11). Table 4 reveals that D-DQN also achieved the lowest false-positive rate (5.2 percent) while preserving trust in benign nodes.

Table 4. Trust isolation accuracy metrics

Note: D-DQN = Dueling Deep Q-Network; A3C = Asynchronous Advantage Actor-Critic; DSR = Dynamic Source Routing.

Figure 11. Accuracy in malicious node isolation

The comparative study demonstrates that D-DQN provides better, more consistent situational and contextual awareness of secure routing performance in an adversarial modus operandi VANET environment. It outperformed A3C and DSR across all major metrics: PDR, latency, trust convergence, and energy optimization. Although A3C was practical, it experienced instability at high threat levels. DSR was not the most resilient one because it lacked adaptive or trust-driven mechanisms. The results support the compatibility of DRL with trust assessment as a helpful way of protecting real-time communication in intelligent transport systems.

5.8 Policy reward accumulation

The reward patterns in Figure 12 show that D-DQN converged quickly and had a smoother reward curve. A3C was also not as consistent because it did not update asynchronously. DSR, which lacks policy learning, exhibited no reward optimization.

Figure 12. Cumulative reward per episode

5.9 Analysis of the results

The comparative study demonstrates that D-DQN provides better, more consistent situational and contextual awareness of secure routing performance in an adversarial modus operandi VANET environment. The results support the compatibility of DRL with trust assessment as a helpful approach to protecting real-time communication in intelligent transport systems, in which D-DQN outperformed A3C and Q-learning across all major metrics: PDR and latency, trust convergence, and energy optimization.

D-DQN is more effective than A3C or tabular DSR due to its ability to learn a stable value function with less overestimation and strong generalization in the high-dimensional and quickly changing VANET states, compared to A3C, which is more susceptible to training variance, and tabular DSR, which cannot be applied to non-discrete state spaces. Although A3C was practical, it experienced instability under high threat. DSR was not the most resilient one because it lacked adaptive or trust-driven mechanisms.