.png){kind=icon}
.png){kind=icon}
Return
Resilient Fully-distributed Reinforcement Learning for UAV Swarms against General Byzantine Attacks
.png){kind=icon}
PRE
AI
Y
J
T
DOI：10.1016/j.jfranklin.2026.108732.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Unmanned aerial vehicles (UAVs) powered via decentralized Multi-Agent Reinforcement Learning (MARL) offer strong flexibility and scalability, but are inherently vulnerable to General Byzantine Attacks (GBAs) that can severely disrupt system performance. Existing studies predominantly focus on maintaining task performance under Byzantine Edge Attacks (BEAs), which involve malicious information propagates through communication links, and largely overlooking Byzantine Node Attacks (BNAs), where compromised UAVs manipulate their policies or observations to mislead others. BNAs pose persistent and systemic risks, and most existing defenses either neglect them or tightly couple attack mitigation with specific task designs, limiting adaptability and scalability. To address these challenges, we propose H-MARL, a hierarchical MARL framework inspired by digital twin technology. H-MARL explicitly decouples system tasks from defense strategies and separates GBAs into BEAs and BNAs for targeted mitigation. It consists of three layers: (i) the Twin Layer (TL), which performs resilient discrete path planning using a projection-based Byzantine-resilient actor-critic (BR-AC) algorithm to counter BEAs without centralized coordination; (ii) the Scheduling Layer (SL), which refines the discrete paths into smooth reference trajectories with higher-order derivatives through an Obstacle-aware Bézier Optimization (OABO) algorithm; and (iii) the Cyber-Physical Layer (CPL), where each UAV employs a decentralized adaptive controller with recursive backstepping and compensation mechanisms to suppress unbounded BNAs during trajectory tracking. Theoretical analysis and experimental evaluations demonstrate that H-MARL achieves bounded consensus under BEAs, effectively mitigates malicious behaviors caused by BNAs, and enhances both the robustness and scalability of MARL-based UAV swarms operating in adversarial environments.
Keywords:
UAV swarms
Multi-Agent Reinforcement Learning
Byzantine Attacks
Digital Twin
Resilient Control
Journal
J
IF:
4.2
Papers:
812
Citations:
0
Organization
No organization information available
Cited Papers