Libertas: Privacy-Preserving Collaborative Computation for Decentralised Personal Data Stores

Rui Zhao, Naman Goel, Nitin Agrawal, Jun Zhao, Jake M. L. Stein, Wael S. Albayaydh, Ruben Verborgh, Reuben Binns, Tim Berners-Lee, Nigel Shadbolt

TL;DR

This paper introduces Libertas, a novel architecture that enables privacy-preserving collaborative computation on decentralized personal data stores (PDS) like Solid. It addresses the challenge of combining secure multi-party computation (MPC) with user autonomy in PDS, allowing individuals to control their data while participating in collective analysis.

What Problem Does This Paper Address?

The increasing value of collective data for scientific, societal, and business insights is undeniable. However, this comes with growing concerns about privacy and data misuse. Decentralized Personal Data Stores (PDS) like Solid offer individuals more control over their data, but performing collective computations on data spread across multiple PDS presents significant privacy challenges. Existing solutions often rely on trusted central parties or struggle with the autonomy and decentralized nature of PDS. This paper tackles the gap in integrating privacy-preserving computation techniques, specifically Secure Multi-Party Computation (MPC), into PDS frameworks while respecting user autonomy and trust preferences.

Methods

Libertas proposes a modular architecture that extends existing PDS frameworks (demonstrated with Solid) to support privacy-preserving collaborative computation. The core innovation lies in its delegated-decentralized MPC model. Instead of each data provider directly participating in MPC, data providers delegate their data to "encryption agents" who then create secret shares of the data. These secret shares are then sent to "computation agents" (players) who perform the MPC. This separation of data providers from computation players is key.

Libertas also introduces an individual-based, user-centric trust model. Data providers define their trust preferences for encryption and computation agents in preference files. The MPC App uses these preferences to dynamically select trusted agents, ensuring that computation facilities are chosen based on individual user trust, not a platform's assumption. This contrasts with the "omniscient" view of security often found in existing MPC literature.

For privacy, Libertas leverages MPC for input confidentiality and suggests the integration of Differential Privacy (DP) for output privacy, enabling the generation of synthetic datasets.

Datasets

Not discussed in the paper. The paper focuses on a framework and its evaluation using simulated data for two use cases.

Key Insights and Ideas

  • Delegated-MPC is Crucial for PDS Scalability: The direct-decentralized MPC model, where every data provider is a computation player, scales poorly with the number of providers. The delegated model, by separating data providers from players, offers significantly better scalability.
  • User-Centric Trust is a Paradigm Shift: Moving from an "omniscient" security model (where a platform dictates trust) to an individual-based, user-centric trust model is essential for PDS adoption. Users should define who they trust for computation, not the other way around.
  • Separation of Duty Enhances Flexibility: The distinct roles of Data Providers, Apps, Encryption Agents, and Computation Agents allow for reusable agent services and flexible delegation of computation tasks, even to third parties like worker unions.
  • Real-World Power Dynamics Can Foster Trust: The paper suggests that existing social structures and power dynamics (e.g., worker unions, data intermediaries) can naturally lead to the establishment of trust relationships required for Libertas's adoption.
  • Synthetic Data Generation is a Powerful Use Case: Combining MPC with Differential Privacy within Libertas opens up new possibilities for generating privacy-preserving synthetic datasets from sensitive decentralized data.

Key Results

  • Scalability: Benchmarks show that delegated-decentralized MPC scales linearly with the number of data providers, while direct-decentralized MPC scales polynomially. This is a critical finding for practical deployment in large-scale PDS environments.
  • Performance: While computation costs increase with complexity (multiplication vs. sum) and stronger security protocols, the delegated model remains more efficient for a large number of data providers.
  • Use Case Validation:
    • Gig Worker Empowerment: Libertas demonstrated its ability to facilitate collective analysis of gig worker data for purposes like wage negotiation, without relying on a central platform and respecting user autonomy.
  • Synthetic Data Generation: The framework successfully generated differentially private synthetic data, showcasing its potential for data sharing and analysis while protecting privacy.
  • Setup Time Dominance: For simpler computations, the "setup time" (encryption agent preparation and connection establishment) becomes a significant factor as the number of data providers increases, highlighting an area for future optimization.

Limitations and Future Work

  • Complex Governance: Decentralized systems face challenges in ensuring accountability and responsibility, which is identified as a key area for future research.
  • Trust Management: While Libertas allows users to express trust, it doesn't provide advanced mechanisms for managing or dynamically updating trust preferences (e.g., tiers of trust, dynamic preferences based on context).
  • Performance Optimization: Further work is needed to optimize performance, especially for real-time processing and in WAN settings. This includes exploring new MPC protocols, refining code, and optimizing data transmission.
  • Broader Contexts: Evaluating Libertas with other data modalities (image, streaming) and in more diverse real-world collaboration scenarios is suggested.
  • User Training and Studies: User training and real-life collaboration studies are crucial for adoption and for demonstrating societal benefits.

Summary updated

BibTeX
@article{zhao2025libertas,
  author = {Rui Zhao and Naman Goel and Nitin Agrawal and Jun Zhao and Jake M. L. Stein and Wael S. Albayaydh and Ruben Verborgh and Reuben Binns and Tim Berners-Lee and Nigel Shadbolt},
  title = {{Libertas: Privacy-Preserving Collaborative Computation for Decentralised Personal Data Stores}},
  journal = {Proc. ACM Hum. Comput. Interact.},
  year = {2025},
  doi = {10.1145/3757490},
}
Figure