Privacy-Preserving AI Inference
Multi-layer client-side entity extraction with cascading span tracking, automatic sensitivity-based tier routing, streaming response reinjection, and information-theoretically secure entity map protection via XorIDA/GF(2) threshold sharing.
Multi-Layer Cascading Entity Extraction
Four-layer cascading extraction pipeline with span tracking. Each layer operates only on uncovered character spans. L1 regex with checksum validation, L2 schema variants with coreference tracking, L3 NER on remaining spans, L4 optional local LLM. Produces typed semantic placeholders and entity map.
Privacy-Preserving Proxy Inference
Client extracts entities, produces redacted prompt with typed semantic placeholders. Only redacted prompt sent to remote provider. Response reinjected with original values on client. Entity map NEVER leaves client device.
Entity Map Protection via XorIDA Threshold Sharing
Serialize entity map, pad with PKCS#7, generate HMAC-SHA256 integrity tag, split via XorIDA/GF(2) into N shares. K shares reconstruct, fewer than K reveal zero information. Shares distributed across separate storage locations. HMAC verified before any processing.
Privacy-Preserving AI Inference System
Client computing device with cascading multi-layer extraction pipeline, entity map in local memory, sensitivity classification, inference tier routing, reinjection engine, and optional entity map protection via threshold sharing over GF(2).
Fourth Layer: Local LLM Extraction
Optional fourth extraction layer activated when average confidence falls below threshold or deep scan enabled. Local LLM receives redacted text + known entity list, returns pronouns and oblique references linked to entities.
Numeric Confidence Scoring
Every entity scored 0.0-1.0. L1 checksum-validated: ≥0.99, L1 pattern-only: ≥0.95, L2 exact: ≥0.95, L2 variant: ≥0.85, L2 coreference: ≥0.70, L3 NER: by entity type.
TEE Inference Subsystem
Nonce-based attestation, enclave hash + vendor ID verification, ephemeral ECDH key exchange, HKDF-SHA256 session key, AES-256-GCM encrypted inference with counter-mode IV. Session key reused within configurable TTL.
MPC Inference with Vendor Diversity
MPC nodes from ≥2 different hardware vendors, parallel attestation, share distribution via XorIDA, binary-to-arithmetic conversion at node boundaries, collaborative computation, response reconstruction from output shares.
Automatic Sensitivity-Based Tier Routing
Three-tier classification: Maximum (classified/weapons/biometric → local only), High (PII/medical/legal → redact+proxy), Standard (no sensitive patterns → split via XorIDA K-of-N). Pattern-based, stateless, sub-millisecond.
Streaming Reinjection State Machine
Processes SSE chunks from inference provider. Buffer scans for complete placeholder patterns, replaces with original values. Partial placeholders retained for next chunk. Stream termination flushes buffer. Post-stream leak detection scans for original values.
Five-Stage Evasion-Resistant Normalization
Five sequential stages entirely on client: (1) strip zero-width Unicode, (2) NFC normalization, (3) decode HTML entities, (4) replace homoglyphs (Cyrillic→Latin, fullwidth→ASCII), (5) collapse space-padded sequences.
Non-Transitory Computer-Readable Medium
Complete system claim on CRM: multi-layer extraction pipeline, typed semantic placeholders, sensitivity classification, tier routing, inference response reinjection, optional entity map protection via threshold sharing over GF(2).