Balancing Personalization and Privacy: New Solutions for 3D Avatars in Mobile Metaverses

CoE-EDP, VisionRICoE-EDP, VisionRI | Updated: 21-06-2024 13:08 IST | Created: 21-06-2024 13:08 IST
Balancing Personalization and Privacy: New Solutions for 3D Avatars in Mobile Metaverses
Rrepresentative image

A study from the Guangdong University of Technology, Singapore University of Technology and Design, and Zhejiang University Hangzhou, China addresses privacy concerns related to using 3D avatars in mobile social metaverses, driven by advancements in spatial computing, extended reality, and AI-generated content. These virtual spaces allow users, termed Social Metaverse Users (SMUs), to interact using personalized 3D avatars. However, generating these avatars from real faces raises significant privacy risks, such as identity leakage. To tackle these privacy issues, the researchers propose a new framework for constructing personalized 3D avatars. This framework leverages a two-layer network model allowing SMUs to create avatars that preserve their privacy. Key elements of this framework include avatar pseudonyms, which are temporary, credible identifiers protecting both the profile and digital identity of the avatars. The authors introduce a novel metric, Privacy of Personalized Avatars (PoPA), designed to evaluate the effectiveness of these pseudonyms in preserving privacy. Additionally, the framework includes pseudonym distribution optimization, modeled as a Stackelberg game, where deep reinforcement learning (DRL) is used to learn equilibrium strategies under incomplete information conditions.

Merging Mobile Edge Computing with Social Networks

The system model integrates mobile edge computing with social networks in metaverses, facilitating immersive social activities like virtual tours and meetings. The 3D avatar construction framework consists of two layers. The first layer focuses on profile image generation, utilizing a generative model (e.g., Stable Diffusion) to create diverse styles of avatar images from text or image prompts. The second layer involves 3D face generation, allowing SMUs to select different models for 3D shape regression and texture map generation to construct personalized 3D avatars. The pseudonym framework minimizes communication overhead by allocating batches of pseudonyms to SMUs. These pseudonyms combine avatar attributes and random parts to ensure both profile anonymity and digital identity privacy, helping to prevent persistent tracking by malicious entities.

Strategic Pseudonym Distribution Using Game Theory

The privacy-preserving performance is modeled using a Stackelberg game, with the Local Authority (LA) as the leader and SMUs as followers. The LA determines the price of pseudonyms, while SMUs adjust their pseudonym demands accordingly. The game's equilibrium ensures an optimal pseudonym distribution strategy, balancing privacy and resource utilization. Given the incomplete information scenario where SMUs do not share sensitive details with the LA, a DRL-based solution is proposed. This transforms the problem into a partially observable Markov decision process, where the LA, acting as an intelligent agent, uses past pseudonym demand data to make optimal pricing decisions.

Ensuring High-Fidelity Avatar Generation and Privacy

The experimental evaluation demonstrates the quality of personalized 3D avatar generation and the performance of the DRL-based pseudonym distribution. The proposed methods generate high-fidelity avatars while maintaining privacy by preventing the use of real faces. Metrics like LPIPS are used to validate the suitability of the method across various scenarios. The DRL-based approach outperforms baseline methods, achieving higher utility and efficient pseudonym distribution even under incomplete information conditions. The framework effectively balances personalized avatar creation and privacy preservation in mobile social metaverses. Future work will focus on enhancing security assessments and incorporating more effective privacy metrics.

Bridging the Physical and Virtual Worlds Securely

In summary, the emergence of mobile social metaverses has led to widespread adoption of avatars as digital representations for SMUs. These avatars, equipped with immersive devices, allow users to engage in virtual spaces. However, existing 3D avatars, typically generated through scanning real faces, raise privacy and security concerns, such as profile identity leakages. To address these concerns, the authors introduce a new framework for personalized 3D avatar construction, leveraging a two-layer network model for privacy preservation. The framework introduces avatar pseudonyms to jointly safeguard profile and digital identity privacy. The Privacy of Personalized Avatars (PoPA) metric evaluates the effectiveness of these pseudonyms. To optimize pseudonym resource allocation, the pseudonym distribution process is modeled as a Stackelberg game, employing Deep Reinforcement Learning (DRL) to learn equilibrium strategies under incomplete information. Simulation results validate the efficacy and feasibility of the proposed schemes for mobile social metaverses.

The continuous advancement in technologies like spatial computing, extended reality, and AI-generated content has catalyzed the development of metaverses. These advancements have accelerated the development of avatars, which serve as digital representations of humans within virtual spaces of metaverses. For example, the Apple Vision Pro headset's virtual avatar application named Persona highlights the growing interest in personalized avatar creation within metaverses. By merging mobile edge computing with social networks within metaverses, the mobile social metaverse paradigm has emerged, catering to individuals’ desires for engaging in immersive social activities anytime and anywhere. Equipped with immersive devices like VR headsets, SMUs can connect to edge servers to access virtual spaces and interact with others through personally crafted avatars.

Despite the promising prospects of 3D avatars in mobile social metaverses, privacy and security challenges should not be overlooked. Since utilizing true faces for 3D avatar construction poses privacy risks, SMUs often opt for animated characters, animals, or other images as profile pictures on social media. Hence, devising a highly decoupled method for 3D avatar generation, enabling SMUs to customize their avatars while ensuring profile anonymity in metaverses, becomes imperative. Given the social nature of mobile social metaverses, avatars may inadvertently expose the privacy of their real-world counterparts during extensive social interactions. Therefore, preserving digital identity anonymity for SMUs is necessary.

Pseudonyms, serving as temporary credible identifiers, have demonstrated their efficacy in concealing true identities in various contexts. Employing pseudonyms to obscure the identities of avatars for SMUs offers a foundational solution for avatar privacy protection. Leveraging edge computing technology, SMUs can request pseudonyms from Local Authorities located at the network edge to safeguard privacy. However, the limited deployment of edge servers in remote areas results in network congestion due to frequent requests for avatar updates and pseudonyms from SMUs, compromising privacy-preserving performance. Hence, devising a feasible pseudonym resource allocation approach is imperative for sustainable pseudonym distribution.

The study presents a personalized 3D avatar construction framework. Key contributions include catering to the diversity of SMUs’ preferences for personalization and privacy protection, presenting a 3D avatar and avatar pseudonym generation framework incorporating a two-layer network model for personalized 3D avatar generation. The generated avatars have corresponding attribute-based pseudonyms distributed to SMUs for effective privacy protection. The novel Privacy of Personalized Avatar (PoPA) metric evaluates the privacy levels of pseudonymous 3D avatars. A Stackelberg game between the LA and SMUs fosters sustainable pseudonym distribution. A Deep Reinforcement Learning (DRL) algorithm solves the game under incomplete information, demonstrating feasibility and efficacy in mobile social metaverses.

  • Devdiscourse
Give Feedback