EIGR-infer: a blockchain-governed framework for decentralised LLM inference via encrypted intent-guided routing, semantic sharding, and four novel privacy mechanisms

Abstract The dominance of centralized cloud providers in large language model (LLM) inference compels users to transmit sensitive queries to third-party servers, sacrificing data privacy and introducing systemic single points of failure. This paper presents EIGR-Infer —a decentralised inference framework that distributes LLM execution across a heterogeneous peer-to-peer (P2P) mesh of consumer-grade edge devices including smartphones, laptops, and local GPU workstations. The framework’s foundational contribution is the encrypted intent-guided routing (EIGR) protocol, which resolves the Finder’s Dilemma: how to direct a query to the semantically appropriate model shard without revealing the query’s content to the routing network. EIGR achieves this through three coordinated mechanisms: on-device semantic intent hashing using HKDF-SHA256 over a local cosine-similarity projection, Kademlia-DHT-based expert shard discovery keyed on the non-invertible hash, and chaos-adaptive onion routing for encrypted delivery. Beyond EIGR, this paper introduces four entirely novel, previously unpublished contributions: (1) DP-SSIM, a Differentially private semantic intent mapping construction providing formal $$(\varepsilon ,\delta )$$ -DP guarantees on routing metadata; (2) the ghost shard protocol (GSP), a deceptive defence mechanism using cryptographically indistinguishable fake shard announcements to deter and detect model-extraction adversaries; (3) federated domain centroid consensus (FDCC), a Byzantine-fault-tolerant gossip protocol for maintaining the shared domain vocabulary without any central authority; and (4) session-adaptive shard re-routing (SASR), a mid-session semantic drift detection and shard migration mechanism for high-quality private multi-turn inference. All seven algorithms are formally specified and analysed. Discrete-event simulation across a 1000-node heterogeneous mesh demonstrates EIGR discovery latency of 47 ms at the 95th percentile, end-to-end inference latency of 460 ms for 512-token prompts, DP-SSIM overhead below 2 ms, ghost shard detection rate of 98.3%, FDCC convergence within 3–4 gossip rounds, and SASR re-routing latency of 31 ms—all compatible with interactive inference targets.