Insights by: Rob Viglione, co-founder and CEO of Horizen Labs
Can you truly rely on your AI to be impartial? A recent study indicates that the situation is more intricate. Regrettably, bias is not merely a flaw—it’s an ingrained issue without sufficient cryptographic safeguards.
A September 2024 research report from Imperial College London illustrates how zero-knowledge proofs (ZKPs) can assist businesses in verifying that their machine learning (ML) models treat all demographic groups fairly while maintaining the confidentiality of model specifics and user data.
Zero-knowledge proofs are cryptographic techniques that allow one party to demonstrate to another that a particular statement is true without divulging any additional information beyond the statement’s validity. However, when we start to define “fairness,” we open a new array of challenges.
Bias in machine learning
In machine learning models, bias can appear in various detrimental ways. For instance, it might lead a credit scoring service to assess an individual differently based on their social circle’s credit scores, which can be inherently unfair. It might also lead AI image creators to inaccurately portray figures like the Pope and Ancient Greeks as belonging to differing racial backgrounds, as was notably seen last year.
Identifying an inequitable machine learning model in practice is relatively straightforward. If the model denies individuals loans or credit based on their social connections, that constitutes discrimination. Similarly, if it alters historical context or treats certain demographics differently to rectify perceived imbalances, that also amounts to discrimination. Both instances erode trust in such systems.
Consider a financial institution leveraging an ML model for loan approvals. A zero-knowledge proof could demonstrate that the model isn’t biased against any group without revealing sensitive customer information or proprietary model details. By employing ZK and ML, banks could validate that they’re not discriminatorily impacting any racial group. This proof could be provided in real-time and continuously, contrasting with the current cumbersome government audits of private information.
The ideal machine learning model? One that neither alters historical contexts nor differentiates individuals based on their backgrounds. AI systems must comply with anti-discrimination regulations such as the American Civil Rights Act of 1964. The challenge lies in embedding that principle into AI systems and ensuring it can be verified.
Zero-knowledge proofs present a technical approach to ensure this compliance.
AI does exhibit bias (but it can be mitigated)
In the realm of machine learning, it’s crucial to ensure that any claims of fairness maintain the confidentiality of the underlying ML models and training datasets. They must safeguard intellectual property and user privacy while allowing sufficient transparency for users to validate that their model is fair.
This is no simple feat. Zero-knowledge proofs provide a viable solution.
Zero-knowledge machine learning (ZKML) employs zero-knowledge proofs to confirm that an ML model performs as advertised. ZKML merges zero-knowledge cryptography with machine learning to create systems capable of verifying AI attributes without exposing the underlying models or data. Moreover, we can utilize this concept alongside ZKPs to identify ML models that treat all individuals equally and fairly.
Recent: Understanding Your Peer — The advantages and drawbacks of Know Your Customer (KYC)
Previously, the application of ZKPs to substantiate AI fairness was significantly limited, as it concentrated on a singular stage of the ML pipeline. This allowed dishonest model providers to create datasets satisfying fairness criteria, even if the models themselves did not measure up. Additionally, using ZKPs often entailed substantial computational challenges and prolonged wait times to generate fairness proofs.
In recent months, advancements in zero-knowledge frameworks have enabled the scalability of ZKPs to assess the comprehensive fairness of models with tens of millions of parameters, securely and provably.
The major inquiry: How can we ascertain whether an AI system is indeed fair?
Let’s explore three prevalent definitions of group fairness: demographic parity, equality of opportunity, and predictive equality.
Demographic parity asserts that the likelihood of a specific prediction remains constant across various groups, such as race or gender. Diversity, equity, and inclusion initiatives frequently utilize this measure to reflect organizational demographics. However, this is not the most suitable fairness metric for ML models, as expecting uniform outcomes across all groups is often impractical.
Equality of opportunity is more intuitive to many people. It ensures that every group has an equal chance for a favorable outcome, presuming equal qualifications. This criterion does not seek to optimize results; rather, it guarantees that all demographics have the same prospects for job opportunities or home loans.
Similarly, predictive equality assesses whether an ML model delivers predictions with comparable accuracy across different demographics, ensuring that no one is unjustly disadvantaged due to their group affiliation.
In both scenarios, the ML model refrains from skewing outcomes for equity’s sake, aiming solely to ensure that no group experiences discrimination. This constitutes a rational and sensible resolution.
Fairness is increasingly becoming a standard
In the past year, various governments, including the US, have made public declarations and established mandates concerning AI fairness and the necessity to shield individuals from machine learning bias. With a new administration in the US, the approach toward AI fairness may shift, reinvigorating focus on equality of opportunity over equity.
As political dynamics evolve, so too do the definitions of fairness in AI, oscillating between equity-driven and opportunity-driven perspectives. We support the development of machine learning models that treat all individuals equally without bias. Zero-knowledge proofs can serve as a robust mechanism to validate that ML models are achieving this without compromising the privacy of data.
While zero-knowledge proofs have encountered numerous scalability hurdles in the past, the technology is now becoming accessible for widespread application. We can leverage ZKPs to ensure the integrity of training data, maintain privacy, and verify that the models we utilize perform as expected.
As machine learning models increasingly influence our daily lives, determining our future employment, college admissions, and mortgage options, it’d be beneficial to have greater assurance that AI systems are fair. However, whether we can reach a consensus on what fairness entails remains an entirely different discussion.
Insights by: Rob Viglione, co-founder and CEO of Horizen Labs.
This article is intended solely for informational purposes and should not be construed as legal or investment advice. The opinions expressed here are those of the author and do not necessarily represent the views or opinions of any particular organization.