Secure and Adaptive Routing in Wireless Sensor Networks Using Meta-Router and TGAT-TrustChain

Wireless Sensor Networks (WSNs) have become indispensable in numerous application domains, ranging from environmental monitoring and industrial automation to defense and healthcare. These networks, comprising spatially distributed sensor nodes with limited energy, computation, and communication capabilities, require robust and adaptive routing mechanisms to ensure data integrity, delivery reliability, and system longevity. However, real-world WSN deployments are often exposed to dynamic environmental conditions, node mobility, and potential security threats, such as packet drops, data falsification, and collusion attacks. Existing routing protocols typically operate under fixed assumptions and static trust evaluations, leading to performance degradation and vulnerability in the presence of non-stationary network behaviors. Conventional trust models used in WSNs fail to capture the evolving nature of node behavior over time as they lack temporal sensitivity and are often disconnected from the routing decision-making process. Furthermore, most routing algorithms optimize for performance metrics like delay or energy without incorporating dynamic trust scores or learning from adversarial behavior. Therefore, the existing limitations accentuate the paramount need for a holistic framework that unifies adaptive routing with a robust, temporally aware trust mechanism underpinned by secure and decentralized validation in process.

Existing routing and trust management approaches in WSNs often treat trust evaluation, routing optimization, and attack resilience as isolated processes. State-of-the-art models, such as trust-aware reinforcement routing and blockchain-enabled frameworks, still rely on static trust aggregation or delayed policy updates, limiting adaptability under dynamic network conditions. Moreover, current solutions rarely integrate temporal trust evolution, cross-layer feedback, and adversarial robustness within a unified design.

To bridge these gaps, this paper introduces Adaptive Reinforcement Trust Blockchain Network (ART-BTNet) — a novel, fully integrated architecture that combines reinforcement meta-learning and blockchain for adaptive and secure routing. The proposed framework contributes five distinct innovations:

1. Meta-Router, enabling rapid adaptation to network changes through reinforcement meta-learning;
2. TGAT-TrustChain, introducing temporal graph attention networks for real-time trust evolution on blockchain;
3. CrossLayer-Shaper, fusing cross-layer metrics into a dynamic reward structure for energy-efficient decisions;
4. GAN-AuditChain, employing generative adversarial testing to evaluate trust resilience under malicious conditions;
5. DualAttn-CoModel, jointly optimizing routing and trust using dual spatial–temporal attention mechanisms.

Together, these modules form an interoperable framework that delivers faster convergence, enhanced energy efficiency, and robust protection against trust-based attacks, outperforming existing routing and trust models.

The design of the proposed ART-BTNet model is grounded on the necessity for a tightly integrated dynamic routing-trust optimization framework for WSNs, where energy and security constraints are both critical and interdependent in process. The architecture of the model brings together reinforcement meta-learning, temporal trust modeling through attention-based graphs, cross-layer reward shaping, and adversarial trust auditing through dual-attention fusions. These components have been selected and integrated not in isolation but with mutual reinforcement in mind in order to ensure considerable complements and high systemic coherence for the model. In essence, as seen in Figure 1, the Meta-Router sits at the heart of the routing engine. It applies the Model-Agnostic Meta-Learning (MAML) paradigm to facilitate extremely rapid policy adaptation for nodes witnessing switching under the influence of either topology changes, buffer congestion, or energy variation in the process. The working mechanism optimizes for fast adaptation across tasks using a second-order gradient-based update. For a particular routing policy characterized by θ and a task-specific loss function Lᵢ(θ), the meta-update rule is derived via Eq. (1).

$\theta \leftarrow \theta-\beta \nabla \theta \sum_{\mathrm{i}} \mathrm{L}_{\mathrm{i}}\left(\theta-\alpha \nabla \theta \mathrm{L}_{\mathrm{i}}(\theta)\right)$     (1)

where, α and β represent the inner and outer learning rates respectively for the process. This equation will ensure that the routing policy can generalize for different network conditions without requiring the process to be relearned each time one occurs. In conjunction with that is the CrossLayer-Shaper which would formulate a composite reward signal R by mere aggregation of the inputs from the MAC, network, and transport layers.

Figure 1. Model architecture of the proposed analysis process

The reward is formulated as the weighted sum of several performance metrics via Eq. (2).

$\begin{gathered}R=w 1 \cdot \int r M A C(t) d t+w 2 \cdot \int r N E T(t) d t+w 3 \cdot \int r T R A N S(t) d t\end{gathered}$     (2)

where, rMAC(t), rNET(t), and rTRANS(t) depict time-dependent penalty or reward functions regarding retransmissions, delays, and ACK failures, respectively, for the case under consideration. However, these weights w1, w2, w3 are being tuned dynamically by means of real-time feedback such that rewards would be well aligned to flexible changes of the network conditions. For trust modeling, the TGAT-TrustChain module uses temporal graph attention networks to learn time-sensitive trust vectors $\mathrm{T}_{\mathrm{t}} \in \mathbb{R}^{\mathrm{d}}$, where trust is a function of both spatial relations and temporal sequence. It is given by updating trust scores expressed via Eq. (3).

$\tau_{\mathrm{t}}=\operatorname{Softmax}\left(\sum \in \mathcal{N}(i) \alpha_{\mathrm{ij}}(t) \cdot W \cdot x_{\mathrm{j}}\left(t-\Delta t_{\mathrm{ij}}\right)\right)$     (3)

where, xⱼ(t − Δtᵢᵉⱼ) = feature of neighbor node ‘j’ at a temporal offset Δtᵢⱼ, W is a learnable weight matrix, and αᵢⱼ(t) are the attention coefficients learned through temporal graph self-attention process.

Eq. (3) ensures that each node’s trust score dynamically reflects temporal shifts (e.g., delayed packets, intermittent failures). The Softmax normalization ensures stable trust scaling among neighbors. This enables real-time adaptation of trust scores as network conditions evolve.

With this equation, one will be able to realize a real-time adaptation of the trust scores by impedance of latency, delayed packet responses, and dynamic behavior shifts. Furthermore, to react to legitimacy, the GAN-AuditChain session introduces a computational adversary which analyzes trust transaction in processes. The generator G(z) synthesizes fake trust events, while the discriminator D(x) aims to distinguish between real and adversarial data samples. Optimization follows the standard GAN minimax framework as expressed via Eq. (4).

$Objective = \stackrel{G}{\min }\left(\underset{\max }{D}\left[\mathbb{E}_{\mathrm{x}} \sim {preal}[\log D(x)]+\mathbb{E} z \sim p z[\log (1-D(G(z)))]\right]\right)$    (4)

This mechanism ensures that trust evaluation remains resilient even in the presence of adversarial or malicious data injection attacks.

This process includes synthetic anomalies into the trust chain, ensuring the resilience of trust scoring mechanisms to the attack scenarios. The discriminator is rendered stronger through backpropagated gradients on turquoise Validated consensus mechanisms through blockchains. The last step of decision-making by the DualAttn-CoModel consists of processing jointly the trust vector $\mathrm{T}_{\mathrm{t}}$ and routing vector ρₜ using a dual multi-head attention layer. The final forwarding score S is given via Eq. (5).

$S=\alpha \cdot f \operatorname{TrustAttn}\left(\tau_{\mathrm{t}}\right)+\beta \cdot f R o u t e A t t n\left(\rho_{\mathrm{t}}\right)$     (5)

This equation performs spatial-temporal fusion — a key novelty of ART-BTNet — ensuring routing paths are both secure and efficient, dynamically adapting to trust fluctuations.

where, fTrustAttn and fRouteAttn correspond to nonlinear attention transformations, while α and β are tunable hyperparameters for balancing trust and routing priorities. This is indeed the first model, which integrates decision metrics spatially and temporally for joint optimization, thus leading to the creation of secure and efficient paths selection sets. An accumulated trust deviation metric is also calculated via Eq. (6) to evaluate node behavior over a time window [t₀, t₁].

$\Delta \tau_{\mathrm{i}}=\left(\frac{1}{t^1-t^0}\right) \int\left|\frac{d \tau_{\mathrm{i}}(t)}{d t}\right| d t$     (6)

This equation specifies that in deriving 'i' for node "i," one can see the rate of change of trust for 'i', thereby revealing unstable or suspicious behavior patterns contrary to expected trust evolution dynamics. Auxiliary loss would be included in the learning objective to penalize extremely varied routing or trust decisions across time, formulated as a regularization term via Eq. (7).

Lreg $=\lambda \cdot \sum\left(\left|\frac{d \rho_{\mathrm{t}}}{d t}\right|^2+\left|\frac{d \tau_{\mathrm{t}}}{d t}\right|^2\right)$     (7)

where, λ is a regularization coefficient controlling smoothness strength. The squared derivatives ensure penalization of abrupt transitions, promoting stable learning and preventing oscillations. This ensures that the system achieves smooth adaptation rather than erratic changes in routing and trust estimation.

Temporal smoothness thus ensures that decisions remain stable in highly dynamic environments minimizing energy wastage and route flapping process. A final form that harmonizes the learning objectives of all modules is therefore defined in a collective global loss function combining routing performance, trust accuracy, and adversarial robustness as expressed via Eq. (8).

Ltotal $=$ LRL + LTrust + LGAN + Lreg      (8)

Each one corresponds to losses computed at the Meta-Router, TGAT-TrustChain, GAN-AuditChain, and regularizer, respectively. This kind of summary formulation reinforces joint optimization of all system objectives rather than isolation sets. To sum up, ART-BTNet is a design that balances the adaptability, energy efficiency, and robustness of trust through the advanced machine learning mechanism combined with process blockchain verification techniques. The interoperability of the Meta-Router, TGAT-TrustChain, CrossLayer-Shaper, GAN-AuditChain, and DualAttn-CoModel reinforces system coherence as it enhances complementary functionality. These eight equations then become the mathematics behind the prototyped model, conceptualizing dynamics of learning, trust, validation, and decision-making through layers of WSN stacks.