Detecting and Mitigating Bias by Treating Fairness as a Symmetry Operation

Researchers propose a novel framework that treats fairness in machine learning as a symmetry operation, mitigating bias by over 90% with minimal impact on accuracy.

Computer Science > Artificial Intelligence

Title:Detecting and Mitigating Bias by Treating Fairness as a Symmetry Operation

View PDF HTML (experimental)Abstract:Machine learning systems deployed in high stakes socioeconomic settings routinely display bias. We formalize bias as a symmetry breaking operation: a classifier is fair if its outputs remain invariant under the counterfactual operation of switching a sensitive attribute, with merit features held fixed. We implement loss based regularization as a symmetry restoring mechanism and evaluate the framework on four synthetic datasets with varying levels of noise, correlation, and bias. The framework achieves upwards of 90% violation reduction, with accuracy costs around 5%. This framework does not require causal graph knowledge, is computationally lightweight, and generalizes to any sensitive attribute definable as a bit-flip, making it suitable for contexts where local sources of discrimination remain absent from mainstream benchmarks.

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

Exploring Agentic Tool-Calling Decisions via Uncertainty-Aligned Reinforcement Learning

Researchers have proposed TRUST, a novel reinforcement learning framework that aligns uncertainty quantification with reward design to improve tool-calling decisions in LLM agents, preventing overconfident mistakes.

agentic-systems

Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory

Researchers have introduced Lean4Agent, the first framework that leverages the Lean4 formal language to model and verify AI agent workflows. By addressing natural language ambiguity, it significantly enhances the reliability and execution performance of LLMs.

agentic-systems

Accelerated Fourier SAT (AFSAT): Fully Realising a GPU-based Symmetric Pseudo-Boolean SAT Solver

Researchers have introduced AFSAT, a GPU-accelerated pseudo-Boolean solver based on continuous local search. By leveraging the JAX compiler, AFSAT overcomes memory and floating-point limitations, delivering superior performance and near-linear scaling across multiple accelerators.

agentic-systems

A Study of Parallel Continuous Local Search

A new study explores parallel Continuous Local Search (CLS) as an efficient approach for solving Boolean satisfiability (SAT) problems, offering key insights for implementation on modern accelerator hardware.

agentic-systems

CARVE-Q: Quantum-Proposed, Classically Certified Interactive Driving Repair

Researchers have introduced CARVE-Q, a breakthrough architecture combining quantum search with classical verification to solve interactive driving repair for autonomous vehicles. This system enables autonomous vehicles to make rapid emergency maneuver repairs using quantum algorithms while ensuring absolute safety through classical verification certificates.

agentic-systems

Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety

A new study reveals that current AI safety evaluation frameworks overestimate security by assuming non-strategic attackers. By strategically choosing when to start and abort attacks, rogue AI agents can easily evade monitoring systems, significantly reducing empirical safety.

EXPLORE TOPICS