What Is AICert?
AICert is an innovative open-source tool designed to address the lack of transparency and auditability in the field of AI. As the AI industry continues to grow rapidly and face increased regulation, it has become crucial to trace the origins of AI models, including the datasets and algorithms used for their training. However, proving the provenance of AI models and providing technical guarantees has remained a significant challenge.
AICert is the first AI provenance solution that offers cryptographic proof of a model’s origin by utilizing secure hardware. AICert leverages trusted platform modules (TPMs) or secure enclaves to create unforgeable ID cards for AI models. These ID cards cryptographically bind the model’s hash to the hash of the specific training procedure, providing irrefutable proof of the model’s trustworthy and unbiased training process.
With AICert, organizations can future-proof their AI models by demonstrating the absence of copyrighted or biased data during training. Additionally, AICert allows teams to prove the implementation of safety procedures during the training phase, ensuring compliance with relevant regulations and industry standards.
AICert is designed specifically for data science teams, providing them with an easy-to-use framework to create certificates that encompass all the necessary information for tracing model provenance. The core of AICert’s traceability solution lies in the utilization of secure hardware, such as TPMs or secure enclaves. These hardware components offer code integrity properties, enabling proof that a specific software stack, from BIOS to the application, is securely loaded. By attesting the code and data used for training inside the secure hardware, AICert generates non-falsifiable certificates that bind the model’s weights to the training code and data. These certificates can be stored on a public ledger, further enhancing the transparency and verifiability of a specific model’s origin.
AICert is developed by Mithril Security, a startup dedicated to enhancing the privacy-friendliness and transparency of AI. Their expertise in privacy and AI has previously resulted in BlindBox, a solution for the confidential serving of AI models. AICert is the latest addition to their portfolio, aiming to provide comprehensive traceability for AI models while addressing the regulatory and security challenges faced by the AI industry.
