Attesting Distributional Properties of Training Data for Machine Learning

The success of machine learning (ML) has been accompanied by increased concerns about its trustworthiness. Several jurisdictions are preparing ML regulatory frameworks. One such concern is ensuring that model training data has desirable distributional properties for certain sensitive attributes. For example, draft regulations indicate that model trainers are required to show that training datasets have specific distributional properties, such as reflecting the diversity of the population. We propose the novel notion of ML property attestation allowing a prover (e.g., model trainer) to demonstrate relevant properties of an ML model to a verifier (e.g., a customer) while preserving the confidentiality of sensitive data. We focus on the attestation of distributional properties of training data without revealing the data. We present an effective hybrid property attestation combining property inference with cryptographic mechanisms.