Sustainable Smart Farm Networks: Enhancing Resilience and Efficiency with Decision Theory-Guided Deep Reinforcement Learning

Dian Chen, Zelin Wan, Dong Sam Ha, Jin-Hee ChoarXiv 2025 · OpenReview

Abstract

Solar sensor-based monitoring systems have become a crucial agricultural innovation, advancing farm management and animal welfare through integrating sensor technology, Internet-of-Things, and edge and cloud computing. However, the resilience of these systems to cyber-attacks and their adaptability to dynamic and constrained energy supplies remain largely unexplored. To address these challenges, we propose a sustainable smart farm network designed to maintain high-quality animal monitoring under various cyber and adversarial threats, as well as fluctuating energy conditions. Our approach utilizes deep reinforcement learning (DRL) to devise optimal policies that maximize both monitoring effectiveness and energy efficiency. To overcome DRL's inherent challenge of slow convergence, we integrate transfer learning (TL) and decision theory (DT) to accelerate the learning process. By incorporating DT-guided strategies, we optimize monitoring quality and energy sustainability, significantly reducing training time while achieving comparable performance rewards. Our experimental results prove that DT-guided DRL outperforms TL-enhanced DRL models, improving system performance and reducing training runtime by 47.5%.

Introduction

Smart farming increasingly relies on Internet-of-Things (IoT) sensors to monitor critical animal health metrics such as temperature, heart rate, and activity levels. However, these networks face a significant "double bind." First, they are often solar-powered and resource-constrained; excessive data transmission leads to battery depletion, while insufficient transmission risks missing critical health events. Second, they are vulnerable to cyber-attacks, including False Data Injection and sophisticated adversarial attacks targeting machine learning models.

The wireless solar sensor-based network model, illustrating how compromised nodes can inject falsified data.

While Deep Reinforcement Learning (DRL) can identify optimal transmission policies in these complex environments, it suffers from a "cold-start" problem—slow initial convergence—and is highly susceptible to manipulation by attackers during its early trial-and-error exploration phase.

Proposed Methodology

To address these challenges, this research introduces a hybrid DT-guided DRL framework. The system operates as a multi-agent environment where DRL agents reside on LoRa gateways to manage data transmission from cattle-worn sensors. The framework integrates three core components:

  1. Decision Theory (DT) as a Mentor: Rather than starting with random actions, the DRL agent is guided by a human-designed DT utility function. This function provides an initial "good enough" strategy based on expected utility, accelerating convergence and providing a stable anchor against adversarial poisoning.
  2. Proximal Policy Optimization (PPO): This DRL algorithm refines the policy over time, moving beyond the static DT model to find global optima in dynamic environments.
  3. Subjective Logic (SL) for Deceptive Data Detection: To ensure data integrity, the system uses Subjective Logic to quantify uncertainty. It calculates a "Projected Distance" between a sensor's reported data and the consensus of other agents to identify and discard falsified information.
The multi-agent DRL environment where agents on gateways observe and act upon the smart farm.

Experimental Setup and Data

The study utilized real-world data from the SmartFarm Innovation Network™ at Virginia Tech, including metrics from EmbediVet implantable devices, halter sensors, and GPS trackers. To test resilience, the authors augmented this with semi-synthetic data simulating seven types of attacks, including Denial-of-Service (DoS) and adversarial machine learning attacks like the Fast Gradient Sign Method (FGSM).

Results and Discussion

The DT-guided DRL approach demonstrated significant improvements over traditional methods:

  • Efficiency: The DT-guided PPO model (DT-PPO) reduced training runtime by 47.5% compared to Transfer Learning (TL)-enhanced models. Unlike TL, which requires massive pre-existing datasets, the DT component provides immediate guidance without prior training data.
  • Performance and Resilience: DT-PPO consistently outperformed standalone PPO and rule-based heuristics in accumulated rewards and monitoring quality. Crucially, under heavy adversarial conditions (such as Neural Trojan attacks), the DT-PPO scheme maintained high monitoring integrity, whereas traditional DRL performance plummeted.
  • Energy-Quality Trade-off: The agent successfully navigated the Pareto front between high monitoring quality and battery longevity, optimizing the energy sustainability of the solar-powered network.

Conclusion

This work presents a robust framework for sustainable smart farming by combining the stability of Decision Theory with the adaptive power of Deep Reinforcement Learning. By significantly reducing training time and enhancing resilience against cyber threats, the DT-guided DRL approach offers a viable path for large-scale, autonomous agricultural monitoring. Future research will focus on real-world physical deployments and scaling the multi-agent coordination to support thousands of sensors across vast geographic areas.

Summary updated

Abstract

Solar sensor-based monitoring systems have become a crucial agricultural innovation, advancing farm management and animal welfare through integrating sensor technology, Internet-of-Things, and edge and cloud computing. However, the resilience of these systems to cyber-attacks and their adaptability to dynamic and constrained energy supplies remain largely unexplored. To address these challenges, we propose a sustainable smart farm network designed to maintain high-quality animal monitoring under various cyber and adversarial threats, as well as fluctuating energy conditions. Our approach utilizes deep reinforcement learning (DRL) to devise optimal policies that maximize both monitoring effectiveness and energy efficiency. To overcome DRL's inherent challenge of slow convergence, we integrate transfer learning (TL) and decision theory (DT) to accelerate the learning process. By incorporating DT-guided strategies, we optimize monitoring quality and energy sustainability, significantly reducing training time while achieving comparable performance rewards. Our experimental results prove that DT-guided DRL outperforms TL-enhanced DRL models, improving system performance and reducing training runtime by 47.5%.
BibTeX
@misc{chen2025sustainable,
  author = {Dian Chen and Zelin Wan and Dong Sam Ha and Jin-Hee Cho},
  title = {{Sustainable Smart Farm Networks: Enhancing Resilience and Efficiency with Decision Theory-Guided Deep Reinforcement Learning}},
  year = {2025},
  eprint = {2505.03721},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.03721},
}
Figure