RMS is built on the foundation of opML (Optimistic Machine Learning), a pioneering approach by ORA that integrates the verifiability, decentralization, and resilience of blockchain technology with computation of AI.
Learn more in the original paper about RMS idea.
The architecture of Resilient Model Services (RMS) leverages blockchain for an onchain service contract, which operates akin to a gRPC protobuf by defining services and their callable methods.
Providers must register and stake on this contract to offer services. When a user requests an AI service, they use a verifiable blockchain light client like Helios to fetch a list of available providers from the blockchain.
The user then selects multiple providers randomly and sends offchain transactions to them, which mimic smart contract function calls but are designated for cloud service execution through a unique chain ID.
Providers perform the computation locally, sign, and return the results to the user. The user verifies these results for consistency. If consistent, the process is complete; if not, discrepancies trigger an on-chain arbitration where providers must prove the correctness of their computations.
Optimistic Machine Learning (opML) is the backbone of RMS, designed to integrate the benefits of blockchain technology with AI computation:
Verifiability: opML leverages blockchain for the correctness of AI computations.
Decentralization: The computational load is distributed across decentralized service providers.
Resilience: By requiring only one honest provider to guarantee correct service, RMS operates under a minority trust assumption, making it resilient against malicious actors.
Transparency: All interactions and results are logged on the blockchain, providing transparency.
N-to-M Model: Unlike centralized services, RMS scales by distributing load across many providers, potentially offering better performance during peak times.
Cost-Effectiveness: Decentralized resources can be cheaper than centralized cloud services, with efficient use of decentralized GPUs or CPUs.
For privacy-sensitive computations, FHE can be integrated, allowing users to send encrypted data for processing, with results verifiable without decrypting the data during arbitration.