MOSAIC: Efficient and Secure Outsourcing of AI Computations
A new protocol called MOSAIC has been developed by researchers to facilitate the secure outsourcing of AI computations from a reliable yet computationally limited client to a powerful but untrusted server. This protocol guarantees that the server does not gain access to either the client's input or the model. The main advancement of MOSAIC lies in its matrix-multiplication masking technique, which accommodates significantly larger matrices compared to previous methods, thus allowing the secure handling of contemporary workloads like large transformer inference. By introducing minimal noise into the multiplication outcomes, MOSAIC optimizes client overhead and achieves runtime speeds that are vastly superior to earlier approaches. Its security is based on decisional LWE and LPN assumptions. The paper can be found on arXiv under identifier 2607.29221.
Key facts
- MOSAIC is a protocol for secure outsourcing of AI computations.
- It allows a trusted but computationally weak client to outsource to an untrusted but powerful server.
- The server learns neither the client's input nor the model.
- Core innovation: a matrix-multiplication masking protocol that scales to larger matrices than prior work.
- Enables safe outsourcing of modern workloads such as large transformer inference.
- Introduces small amounts of noise to the multiplication result, relaxing correctness.
- Achieves optimal asymptotic client overhead and concrete runtimes orders of magnitude faster than prior work.
- Security reduces to decisional LWE and LPN assumptions.
- Addresses error growth with an error-scaling mechanism.
- Paper available on arXiv with identifier 2607.29221.
Entities
Institutions
- arXiv