PETs in Practice: using secure multi-party\ncomputation to enable pay gap analysis in the Greater Boston area

Elea Himmelsbach, Calum Inverarity, Dr Jared Robert Keller, Neil Majithia, Elena Simperl, Claudine Tinsman2024 · DOI

TL;DR

This case study explores how the Boston Women’s Workforce Council (BWWC) utilized secure multi-party computation (MPC) to overcome a decade-long data-sharing deadlock. By implementing a privacy-preserving framework, they enabled Greater Boston businesses to collaboratively analyze gender and racial pay gaps without ever revealing sensitive individual payroll data to each other or a third party.

What Problem Does This Paper Address?

The core challenge was a "trust gap" regarding sensitive data. Since 2013, the Boston Council aimed to conduct granular pay equity studies to address systemic inequalities. However, the project stalled because businesses were unwilling to share raw payroll data due to fears of legal exposure, commercial risk, and potential litigation. Even traditional third-party arbiters were hesitant to steward the data for the same reasons.

The paper examines how Privacy-Enhancing Technologies (PETs)—specifically MPC—can move beyond theoretical research into practical, social-benefit applications by providing a "trustless" way to compute aggregate statistics.

Methods

The researchers at Boston University (BU) developed a web-based MPC protocol based on a secret sharing algorithm.

  • Secret Sharing: Private data is obfuscated and split into "shares" distributed among participants. No single share reveals information, and insights only emerge when shares are pooled according to the cryptographic protocol.
  • Client-Server Model: Unlike traditional peer-to-peer MPC, this used a client-server architecture where BU acted as the service provider. This minimized the computational burden on participating businesses.
  • Asynchronicity: To improve usability, the system allowed participants to join and leave computation sessions at their convenience over a two-week window, rather than requiring everyone to be online simultaneously.
  • Idempotency: The protocol allowed businesses to resubmit data to fix errors without changing the final outcome, providing a necessary guardrail for data integrity.
  • Semi-Honest Protocol: The system used a "passive security" model. It assumes parties follow the protocol but might try to learn information from the results. This was deemed sufficient because the reputational and legal risks of collusion between BU and BWWC acted as a deterrent.

Datasets

The system processes sensitive, granular payroll data contributed by businesses in the Greater Boston area.

  • Comparison Benchmarks: The study uses US Census data as a baseline, though the paper notes that census data is often too abstracted and relies on self-reporting.
  • Output: The resulting longitudinal study (produced bi-annually since 2016) tracks both the raw wage gap (total pay across all positions) and the adjusted wage gap (pay for men and women in the same positions, adjusting for experience and education).

Key Insights and Ideas

  • Socialization vs. Computation: The project took two years from inception to deployment. Surprisingly, the delay wasn't due to technical complexity, but the time required to "socialize" the technology and build trust among non-technical business leaders.
  • The "Familiarity" Requirement: To ensure adoption, BU avoided new software packages. Instead, they built the interface around web browsers and Excel spreadsheets, meeting users where they already work.
  • Benchmarking as an Incentive: Participating businesses receive a private ranking of how they perform relative to their peers. This "value flow" encourages participation by providing actionable internal insights that aren't publicly disclosed.
  • Reputational Guardrails: In real-world PETs, cryptographic security is often supplemented by legal and reputational frameworks. The "semi-honest" model worked here because the stakeholders had too much to lose by colluding.

Key Results

  • Long-term Viability: The system has been successfully used for bi-annual reporting since 2016, proving that MPC is sustainable for longitudinal social research.
  • Policy Impact: The success of this implementation inspired the OPEN Government Data Act (2018) in the US, which requires the federal government to modernize data management practices.
  • Scalability of the Model: The BU team has already adapted this MPC model for other use cases, including the Pacesetters Initiative (tracking spending on minority-owned businesses) and Callisto (a platform for survivors of sexual violence).

Limitations and Future Work

  • Upfront Rigidity: A significant drawback is that the code must be written upfront for each session. If the BWWC wants to add a new indicator mid-session, they cannot; all queries must be agreed upon before the two-week window begins.
  • Customization Barriers: The authors note that it remains difficult to customize cryptographic protocols for a wide set of use cases due to a lack of precedents and "off-the-shelf" templates.
  • Need for Open-Source Tooling: Future work focuses on developing open-source libraries and frameworks to make MPC accessible to less-resourced teams who cannot rely on in-kind support from major research universities like BU.

Summary updated

Abstract

About 1PETs in Practice: using secure multi-party computation to enable pay gap analysis in the Greater Boston area 2 Background 3Challenges in developing an MPC framework for benchmarking 4Accessibility 4Usability 4Explainability 5A comprehensive MPC protocol 5Deployment 5Resourcing, implementation, and maintenance costs 6Value flows and tangible outcomes 7 Conclusion
BibTeX
@misc{himmelsbach2024pets,
  author = {Elea Himmelsbach and Calum Inverarity and Dr Jared Robert Keller and Neil Majithia and Elena Simperl and Claudine Tinsman},
  title = {{PETs in Practice: using secure multi-party\textbackslash\{\}ncomputation to enable pay gap analysis in the Greater Boston area}},
  year = {2024},
  doi = {10.61557/hoqo1911},
}

Loading related content...

Figure