With an alarming rise in the use of machine learning (ML) models in high-stakes societal applications come growing concerns about their fairness and transparency. Instances of biased decision-making have caused an increase in distrust among consumers who are subject to decisions based on these models. The demand for technology that allows public verification of fairness in these models is urgent in order to rebuild consumer trust. However, this process is often obstructed by legal procedures and privacy policies which disallow the disclosure of these models by organizations.
In response to this pressing issue, researchers from Stanford and UCSD have proposed a solution: a system known as FairProof. This couplets a fairness certification algorithm with a cryptographic protocol in order to evaluate the model’s fairness at a specific data point, using the metric of Local Individual Fairness (IF). FairProof can issue personalized certificates to individual customers, which makes it perfectly suited for customer-facing organizations. Additionally, the algorithm is designed to be agnostic to the training pipeline, thereby ensuring its applicability across a broad spectrum of different models and datasets.
FairProof leverages techniques from the robustness literature in order to certify Local IF, while ensuring compatibility with Zero-Knowledge Proofs (ZKPs). This maintains model confidentiality by permitting the verification of statements about private data, such as fairness certificates, without revealing the underlying model weights.
A specialized ZKP protocol is implemented as part of FairProof to increase computational efficiency, strategically reducing the computational overhead through optimization of sub-functionalities and offline calculations. Further, the cryptographical commitments ensure model uniformity, allowing organizations to publicly commit to their chosen model weights while retaining confidentiality. This approach, a popular topic in ML security literature, creates a process that maintains transparency and accountability, while safeguarding sensitive model information.
By combining fairness certification with cryptographic protocols, FairProof addresses concerns about fairness and transparency in ML-based decision-making, fostering trust among consumers and stakeholders.
It’s worth noting that this study has been recognized for its impressive work in creating FairProof, receiving the Best Paper Award at the Privacy-ILR Workshop. The inventors of FairProof urge others to join them in the fight for fairness and transparency in ML models. They also encourage fellow enthusiasts to follow them on a range of social platforms, including Twitter, LinkedIn, Reddit, and their dedicated Telegram and Discord Channels, to stay updated on developments on FairProof and their other notable works.
In conclusion, FairProof is a significant step forward in ensuring fairness in ML models, addressing privacy concerns and creating an environment of increased transparency and trust in high-risk societal applications. The creators hope it will help assuage the increasing consumer distrust arising from the proliferation of ML models in society.
