Confidence-Regulated Generative Diffusion Models for Reliable AI Agent Migration in Vehicular Metaverses

TL;DR

This work tackles reliable AI agent migration in vehicular metaverses by jointly addressing security and performance through a TPB-based trust model and a diffusion-based reinforcement learning framework. By formulating migration as a POMDP and introducing the Confidence-regulated Generative Diffusion Model (CGDM), the approach captures multi-modal decision distributions and stabilizes policy updates with a denoising-consistency loss and an adaptive confidence mechanism. Key contributions include a TPB-based reputation scheme that personalizes RSU trust, bilateral vehicle–RSU collaboration for migration decisions, and a CGDM algorithm that outperforms baseline DRL methods in latency reduction and robustness against cyber-attacks. The results demonstrate improved service continuity and security resilience, highlighting the practical impact for real-time vehicular metaverse deployments and autonomous agent migrations under dynamic network conditions.

Abstract

Vehicular metaverses are an emerging paradigm that merges intelligent transportation systems with virtual spaces, leveraging advanced digital twin and Artificial Intelligence (AI) technologies to seamlessly integrate vehicles, users, and digital environments. In this paradigm, vehicular AI agents are endowed with environment perception, decision-making, and action execution capabilities, enabling real-time processing and analysis of multi-modal data to provide users with customized interactive services. Since vehicular AI agents require substantial resources for real-time decision-making, given vehicle mobility and network dynamics conditions, the AI agents are deployed in RoadSide Units (RSUs) with sufficient resources and dynamically migrated among them. However, AI agent migration requires frequent data exchanges, which may expose vehicular metaverses to potential cyber attacks. To this end, we propose a reliable vehicular AI agent migration framework, achieving reliable dynamic migration and efficient resource scheduling through cooperation between vehicles and RSUs. Additionally, we design a trust evaluation model based on the theory of planned behavior to dynamically quantify the reputation of RSUs, thereby better accommodating the personalized trust preferences of users. We then model the vehicular AI agent migration process as a partially observable markov decision process and develop a Confidence-regulated Generative Diffusion Model (CGDM) to efficiently generate AI agent migration decisions. Numerical results demonstrate that the CGDM algorithm significantly outperforms baseline methods in reducing system latency and enhancing robustness against cyber attacks.

Figures (10)

• Figure 1: A reliable vehicular AI agent migration framework in vehicular metaverses. When a vehicle enters the coverage area of an RSU that does not hold a replica of its AI agent, the vehicle uploads the data required to construct the AI agent. Then, the RSU deploys the AI agent and allocates the computing resources. While driving, the vehicle continuously uploads multi-modal data. After performing environment perception, decision-making, and action execution, the AI agent returns the resulting metaverse services to the vehicle. To ensure an immersive experience for users as vehicles move in real time, replicas of the AI agent are proactively pre-migrated to the next appropriate RSU. Moreover, the RSU can actively implement MTD strategies that dynamically reconfigure network addresses and topology, resulting in a brief link reestablishment.
• Figure 2: The overall architecture of the CGDM algorithm. The actor network $\pi_{\theta}$ generates action by performing $K$ denoising steps on Gaussian noise. The resulting state–action pairs are then evaluated by the double critic networks $Q_\phi$ to guide policy optimization. A denoising consistency loss is computed to measure the confidence to adaptively adjust the optimization objective of the actor. All interaction samples are stored in the experience replay buffer for updates of both the actor and critic networks. To ensure stable training, the target networks are softly updated at each iteration.
• Figure 3: Comparison between CGDM with and without confidence mechanism or denoising consistency term.
• Figure 4: Comparison of normalized test rewards and training time with different denoising steps.
• Figure 5: Comparison of test reward curves of different algorithms.