Causal trust aware federated multi agent reinforcement learning for 6G edge networks

The advent of sixth-generation (6G) wireless networks necessitates a paradigm shift towards intrinsically intelligent, decentralized, and trustworthy architectures to manage the extreme heterogeneity, dynamicity, and scale of edge environments. While federated multi-agent reinforcement learning (Fed-MARL) offers a promising path for privacy-preserving, distributed decision-making, existing frameworks are plagued by a lack of interpretability, vulnerability to untrusted agents, and an inability to generalize from spurious correlations, often leading to unfair resource allocation and suboptimal energy efficiency. To address these critical limitations, we propose CT-Fed MARL, a novel causality-aware and trust-driven federated multi-agent reinforcement learning framework designed for fair, interpretable, and energy-efficient management of 6G edge networks. Our core innovation lies in the synergistic integration of two key mechanisms: (i) a causality-aware representation learning module that leverages structural causal models to discern invariant causal relationships from high-dimensional network data, enabling agents to make robust and generalizable decisions amidst environmental non-stationarities; and (ii) a trust-driven aggregation and incentive mechanism that dynamically evaluates agent contributions based on a multi-faceted trust score incorporating model quality, behavioral reputation, and task relevance to mitigate the impact of Byzantine attacks and free-riders, thereby ensuring fair collaboration and sustainable participation. We rigorously evaluate CT-Fed MARL in a simulated 6G ultra-dense network scenario. Extensive experiments demonstrate that our framework significantly outperforms state-of-the-art Fed-MARL baselines, achieving a 32% improvement in fairness (Jain’s index), a 28% reduction in energy consumption, and a 45% increase in model interpretability, all while maintaining robust performance under various adversarial conditions. Our findings underscore the profound potential of integrating causality and trust into distributed learning, paving the way for truly autonomous, resilient, and human-centric 6G edge intelligence.