From Guidelines to Practice: Integrating Techniques in Development Platforms to Achieve Trustworthy AI

Developers are confronted with a plethora of guidelines to address trustworthy AI (TAI). However, guidelines remain mainly abstract and implications for the development process are often difficult to address technically in a holistic manner across the AI development process of data preprocessing, model development, model evaluation, and inferencing. We synthesize existing research on TAI guidelines and realization of TAI qualities, such as security or fairness, to derive techniques that TAI development platforms need to prioritize and include to guide developers for addressing TAI. Our study reveals 34 techniques for achieving TAI on development platforms along six main technique categories: trustworthy training data, trustworthy model training, tests for trustworthy model evaluation, monitoring and control for trustworthy inferencing, external and internal transparency, and data protection. This study contributes to a better understanding of realizing TAI so that researchers and practitioners can better address TAI requirements and ensure impact of TAI guidelines.