Research map / Learning theory and algorithms / Fairness

The Mainstays of Trustworthy Machine Learning

AAAI 2025 · Learning theory and algorithms › Fairness · approach: Fairness auditing and regularization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

A doctoral research summary surveying the author's work on measuring, mitigating, and maintaining trustworthy ML, covering manipulation of influence functions, fairness auditing flaws, zero-knowledge proofs for verifying model properties, and planned unlearning evaluation.

Problem

Technical trustworthy ML solutions (explainability, privacy, fairness) claimed to meet regulatory requirements are often brittle and fail in practice, and confidential models are hard to verify.

Approach

Organizes trustworthy ML research around three mainstays (measurement, mitigation, maintenance) and proposes using cryptography, notably zero-knowledge proofs and commitments, to verify model properties without revealing weights, while stressing measurement of existing tools' weaknesses.

Key results

• Shows that influence scores can be systematically manipulated by adversaries while remaining computationally feasible (Yadav, Wu, and Chaudhuri 2024).
• Argues that current fairness auditing metrics for NLI models hide unfairness by aggregating across groups and misattribute robustness errors as unfairness (Baldini et al. 2023).
• Proposes a ZKP-based framework where the developer proves a property value with cryptographic commitments, winning a best paper award at the ICLR 2024 PPML workshop (Yadav et al. 2024).

Surprising findings

• Current fairness auditing metrics can hide unfairness by aggregating across groups and can wrongly label robustness errors as unfairness.

Limitations

• Research statement only; the ZKP and unlearning evaluation work are described at a high level without experimental details.
• Unlearning evaluation framework is ongoing and not yet presented.

Experimental setup

Builds on: influence functions third-party auditing zero-knowledge proofs

Closest papers by meaning

• Trustworthy Machine Learning through Data-Specific Indistinguishability (ICML 2025 · similarity 0.85)
• Trustworthy AI Must Account for Interactions (NeurIPS 2026 · similarity 0.84)
• Towards Trustworthy, Efficient, and Scalable Machine Learning (AAAI 2025 · similarity 0.83)
• U-trustworthy Models. Reliability, Competence, and Confidence in Decision-Making (AAAI 2024 · similarity 0.83)
• Verification of Machine Unlearning is Fragile (ICML 2024 · similarity 0.81)
• Position: Bridge the Gaps between Machine Unlearning and AI Regulation (NeurIPS 2025 · similarity 0.81)