Cross-Reality Location Entropy in 6G Vehicular Metaverses

• Cross-reality location entropy is a metric that quantifies privacy risks by integrating physical GPS perturbations and virtual AI agent migrations.
• It leverages Bayesian inference to compute Shannon entropy over candidate AV locations, effectively measuring the uncertainty faced by adversaries.
• Empirical studies using LHDPPO demonstrate a 48.9% privacy improvement with lower adversary confidence and minimal latency and QoS degradation.

Cross-reality location entropy is a quantitative privacy metric developed to assess location privacy in 6G-enabled vehicular metaverse scenarios. It addresses the dual exposure arising from autonomous vehicles (AVs) operating across both physical and virtual domains, specifically accounting for information leakage from reported, perturbed GPS data and the virtual edge server locations where AI agents reside. The metric is grounded in Bayesian inference and information-theoretic principles, serving as both an operational privacy measurement and as an optimization objective for privacy-preserving strategizing in cross-reality environments (Luo et al., 18 Jan 2026).

1. Problem Scope and Motivation

The proliferation of 6G vehicular metaverses introduces complex privacy risks: adversaries can jointly exploit physical location reports (from LBS queries, typically GPS) and virtual interactions (such as the assignment of an AV's AI agent to an edge server) to infer real-world AV trajectories. Traditional location privacy metrics and protections, oriented toward a single domain, do not suffice when adversaries correlate cues across these orthogonal but interlinked spaces. This motivated the formalization of cross-reality location entropy, which enables rigorous characterization of adversarial uncertainty accounting for both physical-world obfuscation and the "virtual trail" left by AI-agent placement (Luo et al., 18 Jan 2026).

2. Formal Definition and Mathematical Framework

At time slot $t$, the cross-reality location entropy $E_m^t$ for AV $m$ is the Shannon entropy of the adversary’s posterior distribution over candidate true locations $\mathcal{P}_m^t$, conditioned jointly on the observed (potentially perturbed) physical location $\tilde{l}_m^t$ and the virtual edge server site $\hat{l}_m^t$. Formally, letting $\breve{l}_m^t$ denote the ground-truth location,

$$E_m^t = -\sum_{p_a \in \mathcal{P}_m^t} \Pr\bigl(\breve{l}_m^t = p_a \,|\, \tilde{l}_m^t = p_b,\, \hat{l}_m^t = p_c\bigr) \log_2 \Pr\bigl(\breve{l}_m^t = p_a \,|\, \tilde{l}_m^t = p_b,\, \hat{l}_m^t = p_c\bigr)$$

The posterior, following the Bayes rule and assuming the physical perturbation and server migration are conditionally independent given the true location, is given by:

$$\Pr(\breve{l}_m^t = p_a \mid \tilde{l}_m^t = p_b,\, \hat{l}_m^t = p_c) = \frac{ \Pr(\tilde{l}_m^t = p_b \mid \breve{l}_m^t = p_a) \,\Pr(\hat{l}_m^t = p_c \mid \breve{l}_m^t = p_a)\,\Pr(\breve{l}_m^t = p_a) } { \sum_{p_x \in \mathcal{P}_m^t} \Pr(\tilde{l}_m^t = p_b \mid l_m^t = p_x) \,\Pr(\hat{l}_m^t = p_c \mid l_m^t = p_x)\,\Pr(l_m^t = p_x) }$$

where:

• $\Pr(\breve{l}_m^t = p_a)$ is the location prior (possibly based on historical data),
• $\Pr(\tilde{l}_m^t = p_b \mid \breve{l}_m^t = p_a)$ is induced by the physical-domain perturbation mechanism (e.g., planar Laplace for geo-indistinguishability),
• $\Pr(\hat{l}_m^t = p_c \mid \breve{l}_m^t = p_a)$ is determined by the privacy-aware migration strategy in the virtual domain.

High cross-reality location entropy corresponds to a diffuse posterior, signifying greater adversarial confusion and thus stronger location privacy.

3. Hybrid Privacy-Protection Actions

To elevate cross-reality location entropy, a hybrid action framework is employed, in which each AV simultaneously applies:

• Continuous location perturbation: Introducing randomization (e.g., Laplace noise) to reported physical coordinates, characterized by actions in polar coordinates $(r_m^t, \theta_m^t)$.
• Discrete agent migration: Strategically relocating the AV’s AI agent among a set of edge servers ($e_m^t \in \mathbb{N} \cup \{S\}$, where $\mathbb{N}$ indexes roadside units and $S$ denotes a LEO satellite).

These actions collectively shape the joint observation space available to adversaries, affecting the resultant entropy.

4. Integration with Utility Optimization

AVs must balance privacy against competing operational demands. The instantaneous utility function at each time $t$ for AV $m$ integrates privacy, latency, and service quality:

$$u_m^t = \omega_E E_m^t - \omega_L L_m^t - \omega_Q Q_m^t$$

where $L_m^t$ is the end-to-end service response latency, $Q_m^t$ is the quality-of-service (QoS) loss (e.g., due to decreased rendering accuracy), and weights $\omega_E$, $\omega_L$, $\omega_Q$ encode application-sensitive priorities. The joint optimization problem seeks trajectories $\{r_m^t, \theta_m^t, e_m^t\}$ for all AVs and time slots that maximize the average system utility subject to feasibility constraints. The solution operates under bounds (such as $0 \le r_m^t \le r_{\max}$ and agent migration constraint $e_m^t \in \mathbb{N} \cup \{S\}$).

5. Empirical Insights and Algorithmic Achievements

Experimental studies on real taxi trajectory datasets and metropolitan RSU topologies reveal that maximizing cross-reality location entropy reduces adversarial inference accuracy. Notably, the LHDPPO (LLM-enhanced Hybrid Diffusion Proximal Policy Optimization) algorithm achieves average entropies near $0.743$ bits—a $48.9\%$ improvement relative to classical geo-indistinguishability which ignores virtual-space side information (Luo et al., 18 Jan 2026). Attacker confidence (the probability of correct localization) sharply declines as entropy increases: from above $90\%$ (low entropy, weak privacy) to below $55\%$ as entropy approaches $0.74$ bits. Simultaneously, joint optimization ensures low service latencies (under $30$ ms) and minimal QoS penalties (below $0.15$ radians of rendering error), evidencing that strong privacy is attainable without substantive service degradation.

Method	Avg. Cross-Reality Entropy (bits)	Adversary Confidence (%)	Latency (ms)	QoS Loss (radians)
LHDPPO	0.743	<55	<30	<0.15
Geo-indistinguish.	≈0.499	>90	>30	>0.15

6. Theoretical and Practical Significance

Cross-reality location entropy operationalizes the fundamental uncertainty experienced by an adversary in holistic, cross-domain environments where physical and virtual features are interdependent. It is directly calculable, adaptable to dynamic mobility and topology configurations, and serves as a tractable reward signal for reinforcement learning-based privacy control approaches. Its adoption represents a shift toward privacy assessment and control that are context-sensitive, multi-source, and rooted in adversarial inference modeling. A plausible implication is that similar cross-reality entropy-based metrics could be generalized to other cyber-physical or metaverse systems wherein multiple correlated information flows shape privacy risk surfaces.

7. Concluding Perspective

Cross-reality location entropy realizes a principled and tractable standard for location privacy in 6G vehicular metaverses characterized by tightly coupled physical and virtual exposure. It provides the foundation for optimizing privacy-preserving actions in concert with latency and QoS guarantees. Its development and empirical validation (Luo et al., 18 Jan 2026) underscore its centrality for next-generation privacy research, reinforcing the need for information-theoretic, adversarially grounded metrics as vehicular and cross-reality infrastructures mature.