🤖 machine learning

Cross-reality location privacy protection in 6G-enabled vehicular metaverses: an LLM-enhanced hybrid generative diffusion model-based approach

This paper proposes a cross-reality location privacy protection framework for 6G-enabled vehicular metaverses that employs an LLM-enhanced hybrid generative diffusion model (LHDPPO) to optimize continuous location perturbation and discrete AI agent migration, effectively balancing privacy preservation with service quality and latency.

Original authors: Xiaofeng Luo, Jiayi He, Jiawen Kang, Ruichen Zhang, Zhaoshui He, Ekram Hossain, Dong In Kim

Published 2026-07-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaofeng Luo, Jiayi He, Jiawen Kang, Ruichen Zhang, Zhaoshui He, Ekram Hossain, Dong In Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a car that isn't just a machine, but a smart companion. It has a "digital twin"—a virtual assistant living in a super-fast, 6G-powered internet world (the "metaverse") that helps it navigate, play games, and make decisions. This is the future of vehicular metaverses. But here's the catch: your real car is moving on the road, while its digital twin is hopping between different servers in the cloud to keep up with you. If a sneaky hacker watches where your real car asks for directions and where your digital twin is living, they can put two and two together to figure out exactly where you are, even if you try to hide. This is called a cross-reality location privacy risk.

To stop this, scientists usually try to either blur your real location (like wearing a foggy mask) or hide your digital twin's location (like moving your avatar to a different server). But doing both at the same time is tricky. If you blur your location too much, your navigation gets useless. If you move your digital twin too far, your games lag. The big question is: How do you balance staying hidden, staying fast, and staying accurate all at once? This is the puzzle a team of researchers from China, Singapore, Canada, and South Korea decided to solve.

The Paper's Big Idea: A Smart, Double-Acting Shield

The researchers propose a new system that acts like a master of disguise, using two tricks simultaneously. First, in the real world, the car slightly jitters its reported location (like a magician shifting a coin slightly to the left). Second, in the virtual world, the car's AI assistant jumps to a different server (like a spy changing safe houses). The goal is to make it impossible for an enemy to connect the dots between the real car and the virtual assistant.

To measure how well this works, they invented a new "privacy score" called cross-reality location entropy. Think of this as a measure of "confusion." The higher the score, the more confused the hacker is about where the car actually is. The researchers wanted to find the perfect mix of "jitter" and "jumping" that keeps the hacker confused without making the car's services slow or inaccurate.

The Problem: It's Too Hard for Humans to Solve

The math behind finding this perfect mix is incredibly complicated. It's a "mixed-integer" problem, which is a fancy way of saying it involves both continuous numbers (how far to jitter) and discrete choices (which server to jump to). Trying to solve this with standard math is like trying to find a needle in a haystack while the haystack is on fire and moving.

The paper argues that old methods, like simple random hiding or just using standard computer learning, aren't good enough. They either slow down the car too much or fail to hide the location effectively when hackers are smart. The authors explicitly rule out the idea that just hiding the identity (anonymity) is enough; you have to actively scramble the location data in both worlds.

The Solution: An AI Coach and a Diffusion Artist

To crack this tough code, the team built a new algorithm called LHDPPO (LLM-enhanced Hybrid Diffusion Proximal Policy Optimization). Let's break down this mouthful into a fun story:

The AI Coach (The LLM): Imagine a brilliant coach who doesn't just give you a score, but explains why you did well or poorly. The researchers used a Large Language Model (LLM) to act as this coach. Instead of humans manually writing the rules for what "good" looks like, they asked the LLM to design the reward system. The LLM looked at the complex situation and suggested new, smarter ways to score the car's actions, noticing hidden patterns that humans might miss. It's like having a coach who reads the rulebook and invents a better game strategy on the fly.
The Diffusion Artist (The GDM): Once the coach gives the rules, the car needs to learn how to move. The team used a Generative Diffusion Model (GDM). Imagine an artist who starts with a canvas full of static noise (random chaos) and slowly, step-by-step, removes the noise to reveal a perfect picture. In this case, the "picture" is the perfect decision on where to jitter and where to jump. The model starts with random guesses and "denoises" them into the best possible strategy. They used two of these artists working together: one for the real-world jitter and one for the virtual-world jump.

What They Found

The researchers tested their new system using real data from taxi movements in Shanghai and real base station locations across China. They simulated a 6G world with cars, satellites, and hackers.

Better Privacy: Their new method made the "confusion score" (entropy) much higher than any other method they tested. In their simulations, it improved privacy protection by nearly 50% compared to some older methods.
Faster and Smarter: Even though they were hiding the location better, the cars didn't suffer from slow internet. In fact, the new method reduced the time it took to get a response (latency) by about 26% compared to the next-best method.
The Coach Matters: When they removed the "AI Coach" (the LLM part) and just used the diffusion artists, the system was slower to learn and didn't perform as well. This proved that having the LLM design the reward rules was a key part of the success.

The Verdict

The paper suggests that by combining a smart "coach" (the LLM) that designs the rules and a "noise-removing artist" (the Diffusion Model) that finds the best moves, we can protect our location privacy in the future 6G metaverse without ruining our driving experience.

It's important to note that these results come from simulations using real-world data, not from a live test on actual 6G roads yet. The authors are confident that their approach works in these digital tests and could be a blueprint for future real-world systems, but they acknowledge that real life might bring new challenges, like hackers who know even more about our driving habits. For now, this research offers a promising, playful, and powerful way to keep our digital and physical selves safe from prying eyes.

Technical Summary: Cross-Reality Location Privacy Protection in 6G-Enabled Vehicular Metaverses

1. Problem Statement

The emergence of 6G-enabled vehicular metaverses allows Autonomous Vehicles (AVs) to operate across integrated physical and virtual spaces. In this paradigm, AVs deploy Large AI Model-powered agents on edge servers (Roadside Units or Low-Earth Orbit satellites) to support intelligent decision-making and immersive experiences. However, this tight coupling creates a novel "cross-reality" privacy risk. Adversaries can correlate the physical location reported by an AV when requesting Location-Based Services (LBS) with the virtual location of its corresponding AI agent on an edge server. By leveraging prior knowledge of routine mobility patterns, collusive adversaries can launch inference attacks to reconstruct AV trajectories and track users, a threat more severe than traditional single-domain privacy attacks.

Existing solutions face significant limitations:

Anonymization-based schemes often degrade in low-traffic environments and fail to fully conceal locations.
Obfuscation-based schemes typically address only physical space perturbation, neglecting virtual space interactions.
Privacy metrics (e.g., anonymity set size, geo-indistinguishability) are confined to single dimensions (either physical or virtual) and cannot accurately capture leakage risks arising from cross-reality correlations.
Optimization challenges: Balancing location privacy, service latency, and Quality of Service (QoS) involves a complex, non-convex mixed-integer problem that is difficult to solve with traditional convex optimization or standard reinforcement learning, particularly in high-dimensional hybrid action spaces.

2. Methodology

The paper proposes a user-centric framework based on hybrid action decisions and a novel LLM-enhanced Hybrid Diffusion Proximal Policy Optimization (LHDPPO) algorithm.

A. Hybrid Action Mechanism
The framework protects privacy through two simultaneous actions:

Continuous Action (Physical Space): The AV applies location perturbation to its reported coordinates when requesting LBS. Modeled using a planar Laplacian mechanism, the AV selects a perturbation radius ( $r$ ) and angle ( $\theta$ ) to introduce controlled uncertainty, balancing privacy against QoS degradation.
Discrete Action (Virtual Space): The AV performs privacy-aware AI agent migration. Instead of always migrating to the nearest edge server (which reveals the true location), the AV strategically selects an agent server (RSU or Satellite) to disrupt the one-to-one mapping between the vehicle and the server, thereby introducing uncertainty in the virtual domain.

B. Cross-Reality Location Entropy
A new privacy metric, Cross-Reality Location Entropy, is introduced to quantify privacy levels. Unlike traditional metrics, this is derived from a multi-condition Bayesian posterior distribution that jointly considers:

The inferred physical location.
The perturbed reported location.
The virtual location of the AI agent.
Higher entropy indicates greater uncertainty for the adversary and lower privacy leakage risk.

C. Optimization Formulation
The authors formulate a mixed-integer optimization problem (P1) to maximize the long-term utility of AVs. The utility function balances:

Cross-reality location entropy (Privacy).
Service response latency (Communication, migration, and computation delays).
QoS loss (Penalty for excessive location perturbation affecting service accuracy).

D. LHDPPO Algorithm
To solve the complex optimization problem, the authors develop the LHDPPO algorithm, which integrates two advanced components:

LLM-Driven Informative Reward Design: Instead of relying solely on manually crafted rewards, an LLM acts as an autonomous reward designer. It iteratively generates, evaluates, and refines reward functions based on environmental feedback and performance comparisons. This process identifies latent factors (e.g., the angle between vehicle-to-server direction and perturbation direction) that manual designs might miss, leading to more robust policy learning.
Dual-Chain Generative Diffusion Model (GDM) for Policy Exploration: The algorithm employs two parallel diffusion chains to handle the hybrid action space:
- A continuous chain models location perturbation ( $r, \theta$ ) using a Gaussian distribution.
- A discrete chain models agent migration decisions using categorical distributions.
  Both chains utilize a reverse diffusion process to progressively refine actions from noise into feasible solutions, enabling effective exploration in high-dimensional spaces. These chains are optimized using Proximal Policy Optimization (PPO) with Generalized Advantage Estimation (GAE).

3. Key Contributions

Framework: A cross-reality location privacy protection framework that jointly optimizes continuous location perturbation and discrete AI agent migration to mitigate spatiotemporal correlation risks.
Metric: The proposal of Cross-Reality Location Entropy, a metric derived from Bayesian inference that effectively quantifies privacy under hybrid action strategies across both physical and virtual domains.
Algorithm: The development of LHDPPO, a novel algorithm that combines LLM-driven reward refinement with dual-chain GDM-based policy exploration. This approach decouples hybrid actions and leverages large AI models to automatically design informative reward functions, improving convergence and robustness.
Validation: Extensive experiments using real-world taxi mobility traces (Shanghai) and base station datasets demonstrate the framework's efficacy.

4. Experimental Results

The proposed framework was evaluated against five baselines: Geo-Indistinguishability (Geo-I), Hybrid PPO (H-PPO), GDMs-assisted Hybrid PPO (GHPPO), Hybrid-GDM (HGDM), and PPO-relaxed.

Convergence: LHDPPO achieved higher and more stable rewards compared to baselines, converging faster and avoiding the oscillations seen in other methods.
Privacy Performance: LHDPPO achieved the highest Cross-Reality Location Entropy (0.740), representing a 48.9% improvement over Geo-I and a 28.4% improvement in overall AV utility compared to the strongest baseline (PPO-relaxed).
Latency and QoS: The algorithm reduced service response latency by 26.9% compared to PPO-relaxed. While QoS loss was slightly higher than some baselines, it remained within acceptable bounds given the significant gains in privacy and overall utility.
Sensitivity Analysis: The framework maintained superior performance even under varying adversary knowledge accuracy, backhaul latency coefficients, and data upload sizes.
Ablation Study: Comparing LHDPPO with a version without the LLM reward module (HDPPO) showed that the LLM-driven reward design improved AV utility by 21.5% and reduced convergence steps by 14.6%, confirming the value of automated reward refinement.

5. Significance and Claims

The paper claims that the proposed framework offers a practical solution for preserving location privacy in the emerging 6G-enabled vehicular metaverse, where traditional single-domain approaches are insufficient. By explicitly modeling the coupling between physical vehicle states and virtual AI agent deployments, the framework effectively mitigates inference attacks that exploit cross-reality correlations.

The authors emphasize that the integration of Large Language Models (LLMs) for reward design and Generative Diffusion Models (GDMs) for policy exploration represents a significant advancement in handling complex, high-dimensional hybrid decision problems in dynamic environments. The results suggest that this approach can maintain strong user immersion (low latency, acceptable QoS) while significantly enhancing privacy protection, making it suitable for large-scale deployment in future 6G vehicular networks.

The paper concludes by noting that future work will investigate adaptive privacy protection under heterogeneous AV mobility patterns and against adversaries with more diverse background knowledge.