When Feasibility of Fairness Audits Relies on Willingness to Share Data: Examining User Acceptance of Multi-Party Computation Protocols for Fairness Monitoring

Paper Title

When Feasibility of Fairness Audits Relies on Willingness to Share Data: Examining User Acceptance of Multi-Party Computation Protocols for Fairness Monitoring

Publication Info

• Topic area: User acceptance of privacy-preserving fairness monitoring protocols in algorithmic systems.
• Keywords: fairness monitoring, multi-party computation, privacy-preserving protocols, user acceptance, GDPR, EU AI Act, data donation, algorithmic fairness, privacy calculus, human-computer interaction.

Background and Problem

• Problem / challenge: Fairness monitoring for high-risk AI systems requires sensitive user data, but privacy regulations like GDPR impose strict constraints on data processing. Multi-party computation (MPC) offers a privacy-preserving solution, but its feasibility depends on user willingness to share data, which remains underexplored.
• Significance: Ensuring fairness in algorithmic systems, particularly in high-stakes domains like hiring, is critical for compliance with regulations and societal trust. Understanding user acceptance of MPC protocols can improve fairness monitoring while safeguarding privacy.
• Motivation and related work: Prior research has focused on algorithmic fairness techniques and privacy-preserving methods like differential privacy and synthetic data. However, user-centered perspectives on MPC protocol acceptance, particularly under GDPR and the EU AI Act, are lacking. This study addresses this gap by examining user preferences and orientations toward MPC designs.

Solution

• Proposed approach: A human-centered investigation of user acceptance of MPC fairness monitoring protocols through an online survey of 833 European job seekers.
• Novelty:
  1. Analysis of user priorities and preferences for MPC protocol design attributes.
  2. Examination of how fairness and privacy orientations and contexts influence protocol acceptance.
  3. Practical implications for designing and communicating privacy-preserving fairness monitoring protocols.
• Procedure and key techniques:
  • Survey design based on Privacy Calculus Theory, categorizing protocol attributes into benefit-related, risk-related, and mixed attributes.
  • Conjoint analysis and direct attribute ranking to measure revealed and stated attribute importance.
  • Regression analysis to explore correlations between user orientations and protocol acceptance.
  • Recruitment of a diverse participant sample under GDPR jurisdiction.

Results

• Concrete findings:
  • In conjoint tasks, users prioritized benefit-related attributes (monetary incentive: 23.6%, fairness objective: 19.2%) over risk-related ones (data storage: 20.3%, privacy protection mechanism: 15.1%).
  • In direct rankings, users emphasized privacy-related attributes (privacy protection mechanism: 19.1%, collected information type: 17.6%).
  • Distributed data storage increased acceptance by ~10% compared to basic anonymization and encryption.
  • Broader data collection and use were paradoxically associated with higher acceptance.
• Advantage over baselines:
  • Users trusted research centers (76.9%) and NGOs (72.8%) more than commercial companies (56.7%) for data storage.
  • Distributed storage in MPC protocols achieved higher user acceptance than simpler privacy mechanisms.
• Experiments / evaluation:
  • Survey of 833 participants, including conjoint tasks, attribute ranking, and regression analysis.
  • Diverse participant demographics (e.g., 60.4% white, 15.1% Asian, 15.0% Black) to ensure representativeness.
• Limitations and future work:
  • Cultural variability in fairness and privacy perceptions was not addressed.
  • The study focused on hiring contexts; findings may not generalize to other domains.
  • Real-world deployment feedback and qualitative studies are needed to complement quantitative findings.

Summary

This study investigates user acceptance of multi-party computation (MPC) protocols for fairness monitoring under GDPR and the EU AI Act, focusing on privacy-benefit trade-offs. Using a survey of 833 European job seekers, it finds that users prioritize privacy-related attributes in direct evaluations but value benefit-related attributes in decision-making scenarios. Distributed data storage and trusted third parties like research centers enhance protocol acceptance. Demographic differences reveal that historically disadvantaged groups prioritize fairness objectives over privacy concerns. The findings inform the design and communication of privacy-preserving protocols to ensure informed consent and inclusivity in fairness monitoring.