💻 computer science

Trust-Aware Deep Policy Routing for Tactical Vehicle Ad Hoc Networks under Joint Mobility and Routing-Plane Attacks

This paper introduces AIRouteOpt, a trust-aware deep policy-routing framework that leverages Proximal Policy Optimization to dynamically select next hops based on link quality and recursively updated trust scores, significantly improving packet delivery rates in tactical vehicle MANETs under joint mobility and routing-plane attacks compared to traditional trust-blind and threshold-based approaches.

Original authors: Rohan Shinde, Kishor Shinde, Parag Chaudhari, Hrishikesh Mehta

Published 2026-09-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rohan Shinde, Kishor Shinde, Parag Chaudhari, Hrishikesh Mehta

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the chaotic environment of modern warfare, communication often relies on vehicles that move independently, forming a temporary network without any fixed towers or cables. These mobile networks, known as tactical vehicle ad hoc networks, allow tanks, drones, and ground units to share critical data like maps, sensor readings, and commands. However, keeping these connections alive is incredibly difficult because the vehicles are constantly moving, causing the links between them to break and reform every few seconds. To make matters worse, an enemy can capture one of these vehicles and turn it into a spy. This captured vehicle can lie about the best path for data to travel, tricking the network into sending messages to a dead end where they are silently dropped. The challenge for engineers is to create a routing system that can handle the constant motion of the vehicles while simultaneously spotting and ignoring these deceptive, captured nodes.

Researchers have developed a new approach called AIRouteOpt to solve this dual problem. Instead of using rigid rules that either blindly trust every neighbor or cut off anyone who looks suspicious, this system uses a form of artificial intelligence that learns to make decisions on the fly. The system observes three key things for every nearby vehicle: the quality of the wireless connection, a history of how reliably that neighbor has forwarded messages in the past, and a calculated "trust score" that updates continuously. Rather than setting a hard line where a neighbor is either trusted or banned, the system treats trust as a smooth, sliding scale. It learns to weigh the risk of a deceptive neighbor against the need to keep moving data forward, adjusting its choices moment by moment as the battlefield shifts.

The researchers tested this system in a detailed computer simulation that mimicked a battlefield with vehicles moving at speeds between 5 and 30 meters per second. They introduced various types of attacks, including "blackhole" attackers that drop every message they receive and "sinkhole" attackers that lie about being close to the destination to steal traffic. In these simulations, when 20 percent of the vehicles were captured and acting as enemies, the new system delivered 63.0 percent of the messages. In contrast, traditional systems that did not use this trust-aware intelligence delivered only about 46 percent of the messages. This significant improvement shows that the ability to detect and avoid lying neighbors is the most critical factor in keeping the network alive under attack.

Interestingly, the study found that this smart, learning system did not necessarily deliver more messages than the very best version of an older, manually tuned system that used a strict cutoff rule. When the researchers compared the new AI to a well-tuned traditional system, both delivered roughly the same amount of data, around 63 percent. The real advantage of the new system was not in delivering a higher volume of data, but in how it achieved that result. The traditional system required a precise, pre-set threshold to decide when to cut off a neighbor; if this setting was slightly wrong, the system would fail completely. The new system required no such setting. It learned the right balance automatically, avoided the risk of human error in tuning, and used less network bandwidth to do so. This makes it more robust and easier to deploy in real-world scenarios where conditions are unpredictable.

The research also revealed the limits of this technology. The system excels when attackers lie about their location to attract traffic, but it offers little advantage against "grayhole" attackers that are stealthy and only drop a small fraction of messages without lying about their position. In these specific cases, the network performs similarly whether it uses the new trust system or not. Furthermore, the study was conducted entirely through computer simulation. While the results are statistically significant and the mathematical theory behind the system is sound, the system has not yet been tested on physical hardware in the field. The researchers note that future work will involve testing the system on actual radio equipment and developing defenses against more complex, adaptive enemies.

Ultimately, the paper demonstrates that in a mobile network under attack, the ability to sense trust is more important than the specific algorithm used to make decisions. The new system proves that a network can learn to navigate a hostile environment by continuously evaluating the reliability of its neighbors, rather than relying on static rules. It offers a path forward for military communications that is resilient to deception, adaptable to movement, and free from the brittle settings that often cause older systems to collapse when faced with unexpected conditions.

=== SUMMARY ===

Technical Summary: Trust-Aware Deep Policy Routing for Tactical Vehicle Ad Hoc Networks

1. Problem Statement

Tactical Vehicle Mobile Ad Hoc Networks (MANETs) face a dual challenge that compromises mission-critical communication:

Mobility-Induced Churn: Vehicles moving at 5–30 m/s cause rapid link failures (1–5 s link lifetimes), forcing reactive protocols to constantly rediscover routes and proactive protocols to flood control traffic.
Routing-Plane Attacks: Captured nodes can manipulate routing metrics to attract traffic and then drop packets. Specific threats include:
- Blackhole: Advertises a cheap path (e.g., hop count 1) and drops all traffic.
- Sinkhole: Advertises inflated metrics to attract flows, then drops them.
- Grayhole: Drops a specific fraction of packets while advertising honest metrics (stealthy).
- Sybil: Multiple identities (not evaluated in this study).

Existing solutions fail to address these simultaneously. Geographic routing ignores adversaries; threshold-based trust extensions (e.g., Trust-AODV) assume static topologies and rely on brittle "all-or-nothing" exclusion thresholds; and standard Deep Reinforcement Learning (DRL) routers optimize for throughput in benign settings but cannot distinguish between a congested honest node and a malicious one.

2. Methodology: AIRouteOpt

The authors propose AIRouteOpt, a trust-aware deep policy-routing framework that couples mobility resilience and routing-plane attack resilience into a single decision process.

2.1 System Model

Network: $N$ vehicles in an $A$ km² area using Software-Defined Radios (SDRs) at 2.4 GHz. Links exist based on SNR thresholds with Rayleigh fading and log-normal shadowing.
Threat Model: A fraction $\beta_a$ of nodes are adversarial. Payloads are protected, but routing metrics are unauthenticated.
MDP Formulation:
- State ( $s_t$ ): Per-neighbor features including link quality (SNR, ETX), a recursively updated trust score ( $\tau_{ij}$ ), hop estimates, and a history of the last 8 routing events.
- Action ( $a_t$ ): Select a next-hop neighbor or buffer the packet.
- Reward ( $r_t$ ): A composite function rewarding packet delivery, penalizing latency and deadline misses, and crucially, incorporating a continuous trust-shaping term ( $\lambda_\tau(\tau_{ia_t} - \tau_{ref})$ ). This term encourages selecting higher-trust neighbors without hard exclusion.

2.2 Trust Estimator

AIRouteOpt utilizes a standard Watchdog Exponentially Weighted Moving Average (EWMA) estimator:
$\tau_{ij}(t+1) = (1 - \eta_\tau)\tau_{ij}(t) + \eta_\tau f_{ij}(t)$
where $f_{ij}(t)$ is 1 if the neighbor's onward hop was delivered, and 0 otherwise. The authors tested a "channel-compensated" variant to down-weight failures on weak links but found no significant gain in their ARQ-equipped channel, opting for the standard estimator.

2.3 Policy Learning

The routing policy is learned using Proximal Policy Optimization (PPO).

Architecture: A shared two-layer Multi-Layer Perceptron (MLP) with 64 and 32 hidden units (tanh activation).
Output: A softmax distribution over $K+1$ actions (neighbors + buffer), masked by neighbor availability.
Training: Trained on a curriculum of mobility and attack fractions. The policy is deployed frozen on each vehicle.
Complexity: Inference requires a single forward pass (~0.035 ms per decision, 5191 parameters), making it suitable for SDR-class hardware.

3. Theoretical Contributions

The paper provides three theoretical results tied to empirical measurements:

Trust Concentration (Prop. 1): The trust estimator concentrates around the neighbor's true forwarding probability ( $\beta^*_j$ ). The authors prove a Hoeffding bound showing that blackholes ( $\beta^*=0$ ) and honest nodes ( $\beta^* \approx 1$ ) are easily separated, while stealthy grayholes (e.g., $\beta^*=0.6$ ) remain difficult to distinguish from honest nodes with high confidence.
Monotonic Improvement under Drift (Thm. 1): The authors derive a bound for PPO updates in non-stationary topologies. They show that the improvement guarantee holds provided the policy update is constrained relative to the topology drift rate ( $\xi$ ).
Threshold Containment (Thm. 2): The authors prove that a soft-trust reward function generalizes hard-thresholding. As the trust weight $\lambda_\tau \to \infty$ , the optimal soft policy converges to the behavior of a hard-threshold policy. This explains why a learned policy might match, but not necessarily exceed, a well-tuned heuristic.

4. Key Results

Experiments were conducted in a custom event-driven simulator over 12–30 seeds with a 20% adversary fraction (70% blackhole, 30% sinkhole).

4.1 Performance vs. Baselines

Trust-Awareness is Decisive: Under metric-advertising attacks, trust-aware methods significantly outperform trust-blind ones.
- AIRouteOpt: 63.0% Packet Delivery Ratio (PDR).
- Trust-AODV (Hand-tuned): 63.9% PDR.
- Trust-Blind Baselines: AODV/OLSR (~44.7%), GPSR (56.3%), Q-routing (40.2%), NoTrust-PPO (45.7%).
Statistical Significance: The gain of AIRouteOpt over trust-blind methods is statistically significant ( $p < 0.001$ , Cohen's $d$ up to 2.0).
Comparison to Trust-AODV: AIRouteOpt is statistically tied with the best hand-tuned Trust-AODV. However, AIRouteOpt achieves this without a brittle exclusion threshold, with lower control overhead (10% vs. 14–38%), and generalizes to unseen network sizes.

4.2 Attack Type Sensitivity

Advertising Attacks (Blackhole/Sinkhole): Trust-aware routing is highly effective.
Stealthy Grayhole: Against a grayhole that does not advertise false metrics but drops 40% of packets, trust-awareness offers marginal benefit. AIRouteOpt (68.2%) performs similarly to trust-blind AODV (72.8%) and slightly below Trust-AODV (73.5%). The trust estimator cannot reliably distinguish partial droppers from honest nodes on fading links.

4.3 Robustness and Scalability

Threshold Sensitivity: Trust-AODV performance collapses if the threshold is set too high (excluding honest nodes) or too low. AIRouteOpt is immune to this hyperparameter tuning.
Scalability: Trained on $N=15$ , AIRouteOpt generalizes to $N=60$ , maintaining performance significantly above AODV.
Latency: AIRouteOpt maintains end-to-end latency within the 200 ms deadline across mobility speeds (up to 30 m/s).

5. Significance and Claims

The paper claims that trust-awareness is the decisive design factor for routing under advertising attacks, rather than the specific learning algorithm.

Primary Finding: A learned soft-trust policy (AIRouteOpt) can match the performance ceiling of the best hand-tuned trust heuristic (Trust-AODV) but removes the operational fragility of hard thresholds.
Operational Value: The advantage of AIRouteOpt over Trust-AODV is not a higher delivery ratio in ideal conditions, but rather:
1. Elimination of the critical threshold hyperparameter.
2. Reduced control overhead.
3. Joint optimization of link quality, trust, and buffering.
Limitations: The authors explicitly state that the benefit is confined to metric-advertising attackers. Against stealthy grayholes, the benefit is marginal. Furthermore, AIRouteOpt does not strictly dominate a well-tuned Trust-AODV in terms of raw PDR; its superiority lies in robustness and operational flexibility. The study is simulation-based and has not yet been validated on hardware-in-the-loop or standard simulators like ns-3.

In conclusion, the paper argues that integrating trust as a continuous reward-shaping term within a deep policy framework provides a resilient, threshold-free routing solution for tactical MANETs, effectively neutralizing blackhole and sinkhole attacks while maintaining theoretical guarantees under topology drift.