Trustworthiness Preservation by Copies of Machine Learning Systems

Trustworthiness can degrade when ML models are copied: The paper addresses a critical gap in responsible AI by investigating whether trustworthiness properties are preserved when the same model is trained on different datasets or when different models are trained on the same dataset—practices common in ML development but rarely formally verified.

Four distinct levels of trustworthiness preservation exist: The authors define specific categories (Justifiably, Equally, Weakly, and Almost Trustworthy) to measure how well a copied model maintains the trustworthiness of the original, providing a framework to distinguish between different degrees of preservation rather than treating it as binary.

Trustworthiness preservation can be computationally verified: The paper introduces a calculus that enables developers to check whether copies of ML systems maintain trustworthiness properties by analyzing the partial behavior of the copy against the original, offering a practical computational tool for responsible AI practitioners.

Trustworthiness properties interact in complex ways: The study examines how different trustworthiness notions compose with each other and under logical operations, suggesting that verifying trustworthiness in complex ML systems requires understanding these interactions rather than checking properties in isolation.

This addresses a "hardly investigated" but legitimate concern: The authors highlight that while creating multiple versions of ML systems is standard practice, the question of whether these copies remain trustworthy has been largely overlooked in responsible AI research.