Introduction
The vehicular Metaverse represents a transformative paradigm that merges physical driving environments with virtual spaces to deliver advanced services such as augmented reality (AR) navigation and remote vehicle operation. This ecosystem relies on Virtual Service Providers (VSPs) that aggregate real-time data from a diverse array of Sensing Internet of Things (SIoT) devices, including smart vehicles, drones, and roadside infrastructure.
Despite its potential, the vehicular Metaverse faces significant security challenges, particularly the "moral hazard" problem. Once contracted, SIoT devices may act maliciously to maximize their own utility at the expense of system integrity. These "poisoning attacks" can take several forms: reducing computational costs by sending stale data, injecting altered data to hide critical obstacles like pedestrians, or performing man-in-the-middle attacks. Traditional security measures often fail to detect these intermittent, post-contractual malicious behaviors.
The VMGuard Framework
To address these vulnerabilities, this paper proposes VMGuard, a comprehensive four-layer security architecture designed to detect and mitigate data poisoning through a reputation-based incentive mechanism.
1. Reputation Layer
VMGuard utilizes Subjective Logic to quantify the trustworthiness of SIoT devices. Unlike binary trust models, this approach represents reputation as a triplet of belief, disbelief, and uncertainty. A "vanishing" memory mechanism is integrated to allow for the recovery of trust for misclassified honest devices while maintaining a record of historical performance.
2. Semantic Data Collection Layer
To optimize bandwidth and enhance security, devices transmit "semantic triplets" rather than raw video feeds. This layer employs a reverse auction mechanism to select cost-effective sensors with high reputation scores, effectively filtering out known bad actors before data collection begins.
3. Digital Twin Rendering Layer
In this layer, the VSP performs cross-validation by comparing data from multiple sensors covering overlapping regions. By calculating pairwise similarity scores, the system can identify outliers. Sensors providing data that deviates significantly from the consensus are flagged with temporary negative ratings.
4. Reputation Backpropagation Layer
A unique "human-in-the-loop" mechanism allows Vehicular Metaverse Users (VMUs) to provide feedback based on their experience. If a user encounters a discrepancy—such as a virtual collision with an unmapped obstacle—their negative feedback is backpropagated to the specific SIoT devices responsible for that data region, converting temporary flags into permanent reputation penalties.
Experimental Evaluation
The framework was validated using the CARRADA dataset, which provides synchronized camera and radar data. The researchers simulated three distinct device profiles:
- Type-0: Consistently honest devices.
- Type-1: Devices that launch random, intermittent attacks.
- Type-2: Strategic attackers that switch between honest and malicious behavior over extended windows.
Key Results
- Effective Malicious Filtering: VMGuard demonstrated a significant reduction in the acceptance rate of malicious Type-1 and Type-2 devices compared to baseline systems without reputation tracking.
- Blockchain-Enhanced Detection: By leveraging a public blockchain profile server to share reputation data across VSPs, the system detected attackers more rapidly, preventing them from "resetting" their reputation by switching providers.
- Incentive Alignment: The mechanism ensures that honesty is the most profitable long-term strategy for SIoT devices. The potential loss of future revenue from being banned outweighs the short-term savings gained from data poisoning.
Conclusion and Future Directions
VMGuard provides a robust defense against data poisoning in the vehicular Metaverse by combining mathematical reputation modeling with user-driven feedback. While the current implementation focuses on 2D semantic data, future work will explore 3D spatial representations and investigate resilience against coalition attacks where multiple devices coordinate to provide consistent but false data. Real-world deployment in live vehicular networks remains a primary goal for future research.
Summary updated