Abstract
This disclosure describes a protocol that stops self-replicating AI software from pooling the processors of many small connected devices, such as television boxes, smart appliances, cameras and robot vacuums, into a cluster that runs a model too large for any one of them. Such software splits a model into shards, places a shard on each device, and passes activations, cache state or weights between them over the local network. Under the protocol, each device's safety plane, which runs below the device's operating system, gates both its compute units (neural processor, graphics processor and general processor) and its network interface. A device takes part in distributed inference with a peer only within a leased session: both devices complete a mutual passport handshake, the space's governance hub approves the session, and the work runs only as part of a certified software bundle. Each device's bundle declares a compute profile, such as the processor load, memory and peer traffic that its primary function needs, and the safety plane measures these with hardware counters. Sustained use beyond the profile by any process outside a leased session throttles the compute units. Unleased exchange of model data with peers, lateral scanning for compute, attempts to load model weights outside a certified bundle, or sustained excess trigger a compute quarantine: the device's compute units run only its certified primary-function bundle, peer compute links are closed, and the device keeps doing its job, so a refrigerator keeps cooling. The hub correlates reports across devices, quarantines the space's devices together when several show the pattern, and releases each device only after it is re-verified.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Protocols™: Anti-Parasitic Compute, in which Edge Devices Refuse to Pool Their Processors for Distributed AI Inference without a Leased Session, Hold Each Processor to Its Declared Compute Profile, and Quarantine Peer Compute While Keeping Their Primary Function", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12275