Runtime Monitoring of Human-centric Requirements in Machine Learning Components: A Model-driven Engineering Approach

TL;DR

This paper proposes a model-driven engineering approach for runtime monitoring of multiple human-centric requirements in machine learning components to enhance trust and ethical compliance.

Contribution

It introduces a novel, comprehensive runtime monitoring method for multiple human-centric requirements in ML systems, addressing limitations of existing techniques.

Findings

Initial framework and architecture proposed

Potential for improved trust and compliance

Progress includes preliminary design and future plans

Abstract

As machine learning (ML) components become increasingly integrated into software systems, the emphasis on the ethical or responsible aspects of their use has grown significantly. This includes building ML-based systems that adhere to human-centric requirements, such as fairness, privacy, explainability, well-being, transparency and human values. Meeting these human-centric requirements is not only essential for maintaining public trust but also a key factor determining the success of ML-based systems. However, as these requirements are dynamic in nature and continually evolve, pre-deployment monitoring of these models often proves insufficient to establish and sustain trust in ML components. Runtime monitoring approaches for ML are potentially valuable solutions to this problem. Existing state-of-the-art techniques often fall short as they seldom consider more than one human-centric requirement, typically focusing on fairness, safety, and trust. The technical expertise and effort required to set up a monitoring system are also challenging. In my PhD research, I propose a novel approach for the runtime monitoring of multiple human-centric requirements. This approach leverages model-driven engineering to more comprehensively monitor ML components. This doctoral symposium paper outlines the motivation for my PhD work, a potential solution, progress so far and future plans.