MOSAIC: Masked Outsourcing for Secure AI Computation

The paper proposes MOSAIC, which securely outsources AI computation from a computationally weak but trusted client to an untrusted powerful server, with the client holding both the input and the model and the server learning neither. Its core is a new matrix multiplication masking protocol that scales to matrices far larger than prior work and supports workloads such as large Transformer inference; by relaxing correctness through adding a small amount of noise to multiplication results, it achieves asymptotically optimal client overhead, with running time orders of magnitude faster than prior work, security reduced to decisional LWE and LPN assumptions, and random Hadamard rotations used to suppress error accumulation across layers. On a 70B model, MOSAIC's perplexity is comparable to mainstream quantization methods, and on HumanEval it is on par with full-precision BF16 inference. The authors also present an end-to-end implementation and state that such methods could enable large-scale confidential AI in modern data centers, maintaining a small trusted computing base (TCB) and outsourcing most computation to untrusted accelerators.