DBLOCK-RLB: DAG Blockchain-Secured Authentication with Dual-Agent Deep Reinforcement Learning and Stackelberg Game-Based Load Balancing for Software-Defined Wireless Sensor Networks

2.1 Routing in Software-Defined Wireless Sensor Network

Since the inception of SDWSNs, energy-aware routing has been the dominant paradigm. An energy-aware routing algorithm that was intended to reduce controller-node interaction by aggregating data was introduced [8]. Nevertheless, this method experiences a substantial increase in control overhead during peak traffic periods as a result of disregarding the status of neighbouring nodes. In order to optimise routing paths, SDN architecture was combined with Q-learning [9]. However, the traditional tabular format and sluggish convergence of Q-learning render it impractical for highly dynamic or large-scale network deployments. To manage traffic routing, deep graph reinforcement learning was implemented in conjunction with graph convolution [10]. Although this method has effectively achieved minimal latency in small topologies, its scalability is severely restricted by the absence of inherent load-balancing mechanisms. In the same vein, a genetic mutation technique based on Particle Swarm Optimisation (PSO) was implemented for the purpose of route selection in heterogeneous SDWSNs [11]. This method is characterised by severe instability when high node densities are present, as it exclusively considers energy and distance metrics. The primary baseline for this study is the Evolutionary Strategy-based Routing Algorithm (ESRA) [12]. Despite the commendable network lifetime extension that ESRA achieves through its evolutionary algorithm-based clustering and routing framework, it is still susceptible to security threats that can indirectly increase energy consumption.

2.2 Clustering mechanisms

Clustering effectively reduces routing overhead by consolidating sensor data at designated CHs before it is transmitted to the base station. Random CH rotation was implemented by the pioneering Low-Energy Adaptive Clustering Hierarchy (LEACH) protocol to distribute the energy load [13]. However, it failed to account for residual node energy, resulting in premature node mortality. Particle Swarm Optimisation (PSO) was implemented in order to resolve this issue by identifying CHs according to an energy-center parameter [14]. However, their methodology failed to consider data privacy and node trust. Energy-efficient CH election has been accomplished in multi-hop networks through the use of Type-2 fuzzy logic [15], and a chaotic genetic algorithm was introduced to optimise joint energy and load balancing [16]. However, the latter continues to be susceptible to malignant genetic manipulation attempts. Fuzzy logic was integrated with LEACH-FC to improve routing reliability, despite the fact that its static membership functions restrict adaptability in an environment that is extremely dynamic [17]. By incorporating a circular partitioning technique, dynamic radius modification, and multi-criteria CH selection, DBLOCK-RLB enhances these methodologies.

2.3 Blockchain-based security

The integration of blockchain technology in WSNs has been expedited, with the primary objective of establishing decentralised trust frameworks and tamper-evident audit traces. The basic security requirements are effectively addressed by a lightweight blockchain authentication mechanism proposed [18]. However, it fails to account for mining scalability in resource-constrained nodes. A conventional blockchain architecture was integrated with Red Deer Algorithm-based clustering in order to improve data security [19]. However, their synchronous block validation paradigm incurs processing delays that are incompatible with real-time Internet of Things (IoT) applications. DAG-BTLBR, a DAG ledger infrastructure for trust-aware, load-balanced routing coupled with lightweight authenticated encryption, was introduced to mitigate these latency bottlenecks [20]. However, this framework is still susceptible to sophisticated multi-vector attacks, despite the inclusion of the Blake-256 and XTEA-Chaotic Map cryptographic algorithms. By extending the DAG paradigm and reinforcing authentication rigour, the proposed DBLOCK-RLB architecture addresses these limitations. Compared to conventional Proof of Work (PoW) and Proof of Stake (PoS) mechanisms, it achieves a significantly reduced computational overhead and latency by employing PoAH consensus and CEA-based key generation.

2.4 Load balancing in Software-Defined Networking

When deploying SDWSNs with many controllers, load balancing is crucial for avoiding controller congestion. Although gradient descent was used to build multipath routing to spread network loads [21], their method is runtime-adaptable since it relies on static path selection. Although a blockchain based on DAGs was used for load balancing and route discovery to extend the lifetime of networks [22], their model does not take into consideration changes in traffic dynamics. As a more formal approach to dealing with operational imbalances, the Maximum Load Minimal Allocation (MIMAL) heuristic was created to solve the load balancing problem (LB-SDN, the second baseline in this study) as an Integer Linear Programming (ILP) problem [23]. However, their method is rather slow because slave controllers need sequential verification. By using a Stackelberg game-theoretic trading paradigm, the suggested DBLOCK-RLB design reduces the impact of this bottleneck. This architecture reduces signalling time to a minimum by allowing direct, incentive-aligned bilateral load transactions supervised by a world leader, thus doing away with per-controller sequential verification. Table 1 provides an organised overview of all of these linked studies.

4.1 Overall system architecture

The DBLOCK-RLB framework comprises four hierarchical planes, as shown in Figure 1. All inter-node and controller communications are recorded as immutable transactions in the DAG-based DBLOCK, which supports authentication, flow-rule validation, and load-trading audits. A high-resolution architecture diagram is provided.

1.png

Figure 1. Layered system architecture of the DBLOCK-RLB framework

Figure 1 illustrates the four-plane hierarchical architecture of the DBLOCK-RLB framework. The Data Plane hosts the sensor nodes and CHs; the Switch Plane enforces DAG-validated flow rules; the Control Plane coordinates the global and local controllers via the Stackelberg trading model; and the Application Plane provides cloud-level analytics and data storage.

• Data Plane. Contains N sensor nodes deployed randomly over the sensing field. Nodes are authenticated prior to cluster formation, and CHs are selected using the ATNP algorithm.

• Switch Plane. Comprises OpenFlow-compatible switches with attached edge servers that validate and enforce flow rules stored in the DBLOCK.

• Control Plane. Hosts one global controller (GloL) and K local controllers (LocC1...LocCk), each equipped with a dedicated load balancer. The Stackelberg game model governs inter-controller load trading.

• Application Plane. Provides application-layer services and a cloud server that stores aggregated sensor data and serves analytics workloads.

• Trusted Authority. Shows the two core security operations: CEA key generation.

4.2 Energy consumption model

The first-order radio energy model is adopted to compute node-level energy expenditure during transmission and reception. Let $P l^B$ denote packet length in bytes and dis the Euclidean distance between transmitter and receiver. The transmission energy is given by:

$\begin{gathered}ETx\left(P l^B, dis\right)=P l^B \cdot EFS \cdot dis ^2+P l^B \cdot \text { Eckt } \\ if \, dis \leq diso=P l^B \cdot EMPF \cdot dis^4+P l^B \cdot Eckt \\ if \, dis>diso\end{gathered}$              (1)

where, $P l^B$ denotes packet length (bits), dis is the Euclidean distance (m) between transmitter and receiver, EFS = 11 pJ/bit/m² is the free-space channel energy coefficient, EMPF = 0.0017 pJ/bit/m⁴ is the multipath fading coefficient, dis₀ = √(EFS/EMPF) ≈ 80.5 m is the crossover distance, and Eckt = 48 nJ/bit is the circuit energy consumption. Reception energy:

$\operatorname{ERx}\left(P l^B\right)=P l^B \cdot \operatorname{Eckt}$             (2)

Idle-state energy:

Eidle = Tidle ⋅ $($ Pidle/Vidle $) / 1000$           (3)

where, Tidle = LePS − TTxctrl − TTxdata − TRx is derived from the sensing period LePS (s) and the control transmission time TTxctrl (s), data transmission time TTxdata (s), and reception time TRx (s). All parameter values are listed below. Node Authentication via Camellia Encryption and DAG Blockchain

4.2.1 Registration phase

The Trusted Authority (TA) gets credentials from all sensor nodes (Sn) in the form of an ID, a PUF, a MAC address, and a position. The CEA is used by the TA to make a 256-bit secret key called $S K_2 56$. The plain CRD is XOR-combined with the left and right subkeys $\left(S K_L e^2 56, S K_R i^2 56\right)$ of equal length:

$\begin{gathered}\operatorname{Pla}(C R D)_2 56 \times O R\left(S K 1_b(28) \| S K 2_b(128)\right)=L e_n(128) \| R i_0(128)\end{gathered}$              (4)

For rounds b = 1 to 24, the round transformation applies:

$L e_b=R i_{b-1} X O R F\left(L e_{b-1}, S K_b\right)$            (5)

$R i_b=L e_{b-1}$             (6)

The final secret key is:

$\begin{gathered}S K_C R D(256)=\left(R i_2 4(128)| | L e_2 4(128)\right) \\ X O R\left(S K 3_b(128)| | S K 4_b(128)\right)\end{gathered}$                 (7)

The generated $S K_C R D(256)$ is stored in the DBLOCK as a hashed transaction:

$D B L O C K<-T A\left(H\left(S K_C R D(256)\right)\right)$           (8)

4.2.2 Authentication phase

During network operation, the TA decrypts the node-provided secret using reverse CEA rounds (b = 24 to 1) and extracts the embedded CRD. If the current timestamp CRD matches the registered CRD, the node is authenticated; otherwise, it is revoked:

$\begin{aligned} T A= Authenticated, if \, C R D_n^t==C R D_n^{t-1} \\ \{Revoked, otherwise\} \end{aligned}$             (9)

The PoAH consensus verifies each block in two steps: (i) the current block authenticates the transaction against its predecessor block, and (ii) the authenticated block's trust score is incremented by one unit. Any miner endorsing a false block loses trust and is eventually excluded from the mining pool. This design avoids inverse hash computation, contributing to the energy efficiency reported in Section 5.

Trust Score Management: A trust score τᵢ is maintained by each authenticated node vᵢ, which is initially set to 1.0 during registration. Trust is increased upon successful block validation: τᵢ(t+1) = τᵢ(t) + 0.1. Trust is reduced by a more severe penalty in the event of a failed or fraudulent validation: τᵢ(t+1) = τᵢ(t) - 0.3. A node that is blacklisted and excluded from all subsequent mining and routing operations is promptly excluded if its trust falls below τ_min = 0.3. The rapid isolation of malevolent actors is guaranteed by this asymmetric trust adjustment scheme.

Collusion Resistance: PoAH necessitates that each new block be validated against two parent blocks in the DAG. Consequently, successful collusion necessitates a minimum of three colluding nodes with trust scores exceeding τ_min, resulting in a substantially higher coordination cost than single-endorser PoS schemes. Although formal game-theoretic collusion resistance proofs are beyond the scope of this simulation study, we acknowledge this as a significant direction for future theoretical research. Table 2 provides quantitative consensus comparisons.

Table 2. Quantitative consensus comparison

4.3 Adaptive Threshold-based Network Partitioning

Only authenticated nodes $A u S_N=A u S_1, A u S_2, \ldots, A u S_n$ participate in cluster formation. The SDWSN field is modeled as a circular area partitioned into equal sectors using an adaptive threshold based on node density and inter-node distance:

$\operatorname{Cir}\left(A u S_N\right)=\operatorname{cir}^H * \exp -\left(A u S_n-\operatorname{Clu}_c e n\right)^2 / \operatorname{Clu}_r a d\left(A u S_N\right)$              (10)

where, $cir^H$ is the number of partitions, $Clu_c e n$ is the cluster centroid, and $Clu_r \mathrm{ad}$ is the adaptive radius. During sudden topology changes (node failure, mobility), the partition radius is updated as:

$\mathrm{Cir}_{n e w}^{H n}=\mathrm{Cir}_{o l d}^{H n}+1 /\left(\mathrm{Cir}_{o l d}^{H n}+1\right)$           (11)

This formulation ensures that as network density decreases (node death), sectors expand to maintain coverage continuity. CH selection within each sector is performed by each node broadcasting a composite fitness value V:

$\begin{aligned} V=w_E * E_r e s i d u a l & +w_D *(1 / d i s t)+w_C \\ & * \text { centrality }+w_K \\ & * \text { connectivity }+w_L \\ & * \text { link }{ }_{\text {stability }}\end{aligned}$           (12)

where, w_E, w_D, w_C, w_K, w_L are weight coefficients summing to unity, calibrated empirically. The node with the highest V within its sector is elected as CH.

4.4 Dual Agent-Twin Delayed Deep Deterministic Policy Gradient intelligent routing algorithm

The chosen CH uses the DA-TD3 algorithm to do the routing. This is an actor-critic deep reinforcement learning method in which two cooperative agents operate over the SDWSN state space, as illustrated in Figure 2.

2.png

Figure 2. Dual Agent-Twin Delayed Deep Deterministic Policy Gradient (DA-TD3) dual-agent routing architecture for Software-Defined Wireless Sensor Network (SDWSN)

The attributes and responsibilities of the DA-TD3 agents utilised in the proposed routing mechanism are summarised in Table 3. Agent W1 identifies the optimal forwarding node, while Agent W2 determines the most efficient routing path to accomplish the overall network optimisation goal. The two agents operate collaboratively.

Table 3. Dual Agent-Twin Delayed Deep Deterministic Policy Gradient (DA-TD3) agent attributes and roles

Both agents share two centralized critics $Y_1, Y_2$ and a common actor $theta_I N$. The policy gradient update is:

Throughput, Latency, PDR, Network Lifetime, and Energy Consumption are five popular SDWSN indicators that are used to measure performance. For each measure, the results are shown as the mean plus or minus the standard deviation of 10 different simulation runs.

6.1 Throughput analysis

Throughput T is defined as:

$T=P / t d o t$          (22)

where, tdot is the simulation time that has passed and P is the number of packets that were successfully received. At 100 nodes the primary benchmark configuration DBLOCK-RLB achieves a mean throughput of 2.55 ± 0.1 Mbps, compared to 1.85 ± 0.2 Mbps for ESRA and 1.55 ± 0.3 Mbps for LB-SDN. The upward trend observed in Figure 5 as node count increases reflects the improved spatial reuse and load distribution afforded by finer ATNP partitioning at higher densities. The improvement is statistically significant (paired t-test, p < 0.01). Figure 5 plots throughput against node density for all three methods. DBLOCK-RLB (purple triangle, solid line) is the highest performer at every density tested, reaching 2.55 ± 0.1 Mbps at the 100-node benchmark configuration; ESRA (green circle, dashed line) is the mid-tier performer, reaching 1.85 ± 0.2 Mbps at 100 nodes; and LB-SDN (coral square, dotted line) is the lowest performer, reaching 1.55 ± 0.3 Mbps at 100 nodes. Each series uses a distinct line style (solid / dashed / dotted) and marker shape (triangle / circle / square) so the figure remains interpretable when printed in grayscale.

These 100-node values correspond to the benchmark figures reported below. Error bars represent ±1 SD across 10 simulation runs.

6.2 Latency analysis

End-to-end latency L is computed as:

$L=t_r$ eceived $-t_s$ ent          (23)

At 100 nodes, DBLOCK-RLB achieves 27.1 ± 0.1 ms, which is 10 ms faster than ESRA (37.3 ms) and 24 ms faster than LB-SDN (51.2 ms). This gain is driven by two mechanisms: (i) trading-based load balancing reduces scheduling latency by avoiding sequential controller verification; and (ii) PoAH consensus speeds up block validation by removing the computational burden of inverse hash verification. For all approaches, latency rises almost linearly with node count; however, DBLOCK-RLB's slope is 23% smaller than LB-SDN's, demonstrating the Stackelberg trading model's improved scalability.

Figure 6 illustrates the influence of node density on end-to-end latency in various routing methodologies. In dense network environments, the proposed DA-TD3 method consistently maintains lower latency than baseline techniques as the number of nodes increases, thereby demonstrating its efficacy in reducing communication delays and enhancing routing efficiency.

Figure 6. End-to-end latency comparison across node density

6.3 Packet Delivery Ratio analysis

PDR is defined as:

$P D R=\left(D_r\right.$ eceived $/ G_q$ enerated $) * 100$          (24)

DBLOCK-RLB achieves 91.5 ± 0.1% PDR at 100 nodes, compared to 81.5% for ESRA and 75.5% for LB-SDN. By ensuring that data is routed through nodes with adequate residual energy and stable links, multi-criteria CH selection (Eq. (12)) lowers mid-path failures. Proactive avoidance of lossy links is made possible by the DA-TD3 route selection agent's consideration of packet loss rate as a state feature. When combined, these strategies provide a 10.0 percentage-point advantage over the next best rival.

The PDR for the evaluated routing protocols is compared across various node densities in Figure 7. The results suggest that the proposed DA-TD3 approach maintains superior packet delivery performance as network density increases, underscoring its efficacy in minimising packet loss and ensuring dependable communication.

Figure 7. Packet Delivery Ratio (PDR) comparison across node density

8.1 Performance interpretation

Three complimentary methods working at various network tiers are responsible for the performance improvements mentioned in Section 7. When compared to public-key cryptographic systems, CEA-based authentication lowers the credential verification overhead per authentication round at the node level, which lowers the energy expenditure on security operations per node [36]. By distributing transmission load among nodes with high residual energy and central connectivity at the cluster level, ATNP's multi-criteria CH election avoids the hot-spot depletion that shortens ESRA's network lifetime. LB-SDN's expensive per-slave-controller polling cycle is eliminated at the controller level by the Stackelberg trading model's direct bilateral load transfer, which reduces average scheduling latency by 24 ms. A detailed examination of the DA-TD3 algorithm's contribution to PDR improvement is necessary. In continuous action spaces, classical Q-learning-based routing [37] suffers from Q-value overestimation, which leads the agent to favour congested paths that seem ideal under inflated value estimations. This is directly mitigated by DA-TD3's twin-critic architecture, which generates conservative but reliable Q-value estimations by using at least two independently trained critics. In the non-stationary SDWSN environment, where connection quality and node energy fluctuate constantly, the delayed actor update further stabilises policy learning.

8.2 Scalability analysis

The advantages of DBLOCK-RLB over ESRA increase monotonically with network size, as seen in Table 7. In contrast to ESRA's sequential blockchain validation, the DAG blockchain's parallel transaction processing allows each transaction to be verified by concurrently validating it against two parent transactions, giving it a scalability benefit. The reported latency and energy gains are directly related to the DAG structure's 3.1x faster processing of authentication transactions at 100 nodes compared to a linear chain (measured in block confirmation time). Additionally, the ATNP clustering algorithm shows good scalability: as the number of nodes grows, the adaptive threshold creates finer-grained partitions that keep intra-cluster communication distances roughly constant, avoiding the per-cluster energy growth that would otherwise happen with fixed-partition clustering.

8.3 Limitations

DBLOCK-RLB has a number of drawbacks that restrict how current findings may be interpreted, despite its excellent empirical performance:

• Simulation fidelity: All tests were carried out using idealised channel models in NS-3.26. Multipath fading, shadowing, and interference from nearby networks are examples of real-world propagation effects that might impair performance, especially the link quality estimates that DA-TD3 Agent W1 uses [38, 39].

• Static topology: The deployment of 100 nodes is predicated on static node locations. ATNP's radius adaption mechanism and DA-TD3's state generalisation will be tested in mobile sensor settings, which are prevalent in vehicle IoT and wildlife monitoring applications [40, 41].

• CEA key length: Although the 256-bit key offers robust security, nodes with very limited resources (such as 8-bit microcontrollers with less than 4 KB RAM) may experience non-trivial computational overhead. It is worthwhile to look into lightweight key scheduling variations [42].

• Resilience to controller failure: The existing Stackelberg paradigm presupposes that at least one local controller and the global controller continue to function. There is no evaluation of mechanisms for graceful degradation under partial controller failure [43].

8.4 Ablation study

To isolate the contribution of each DBLOCK-RLB component, we evaluated four ablated variants against the full system at 100 nodes (10 runs each). The ablation study assesses the contribution of individual components of the proposed DA-TD3 framework in a 100-node network scenario, as illustrated in Table 9. The results indicate that each component contributes to the overall performance, with the complete model attaining the highest efficiency, reliability, and network optimisation.

The full DBLOCK-RLB consistently outperforms all ablated variants, confirming that each module contributes meaningfully to the overall performance. Removing DA-TD3 produces the largest throughput degradation (−18.4%), while removing ATNP causes the greatest energy increase (+22.1%), and removing the Stackelberg model leads to the highest latency increase (+19.2%).