Quality Assurance for ML Devices A Risk-Based Approach

As machine learning systems permeate everyday life, the urgent need for safety, fairness, and quality assurance intensifies. Various frameworks exist to guide this process, yet the challenge lies in seamlessly applying these tools in real development scenarios. In our previous work, we introduced a risk assessment tool grounded in the Failure Mode Effect Analysis (FMEA) methodology and some of these existing frameworks. While the tool showed promise in practical industrial settings, it lacked a comprehensive evaluation and actionable strategies to address the risks found. In this paper, we unveil an improved version of this tool, drawing on experiences and insights from diverse case studies in collaboration with an industry partner. The new methodology offers a robust, end-to-end guideline for evaluating machine learning systems based on a comprehensive and practical functionality analysis, risk assessment, and countermeasure planning. Following multiple workshop evaluations, the tool demonstrated its potential to enable safer and more reliable deployment of AI technology, by helping to visualize the complexly correlated issues affecting such systems, and to navigate the tradeoffs when solving them.