University of Florida researchers proposed SANAI, protecting critical neural-network weights against bit-flip attacks with zero additional parity-storage overhead.

University of Florida researchers published a proposal called SANAI for protecting critical neural-network weights against bit-flip attacks. The paper describes a sensitivity-guided, bit-plane-selective scheme, identifying selective weight protection as a central part of the approach.

SANAI uses embedded error-correcting code (ECC) parity, with the abstract excerpt describing keyed embedding across a two-dimensional weight structure. The stated storage result is zero additional parity-storage overhead.

That claim is specifically about parity storage, not proof of zero computational cost, zero latency overhead, or protection against every attack. The supplied excerpt provides no model-by-model accuracy results, attack success rates, runtime measurements, or production deployments, so the update remains a published research proposal with a stated storage property.