About this article

• Audience: Security reviewers, integrators, and operators mapping Ghost Protocol controls to NIST AI risk frameworks
• Goal: Understand how Govern, Map, Measure, and Manage functions apply to on-device detection, face-quality gates, and video-surveillance cyber practices
• Type: Concept

Summary

Ghost Protocol aligns edge RF-DETR precomp detection, optional NIST FATE face-quality checks, and proxy-owned cloud escalation with the NIST AI Risk Management Framework (AI RMF). On-device person counts stay authoritative; higher layers never override precomp metrics. Full RF-DETR parameters live in the model card.

Prerequisites

• Architecture overview — tiers, dataflow, and escalation order
• Edge AI and RF-DETR — pipeline and precomp authority
• ONVIF reference — standard camera and event surfaces
• COTA POC NIST map — Zero Trust tenets mapped to the 2026 bus AI POC

NIST AI RMF mapping

RMF function	NIST intent	Ghost Protocol mapping
Govern	Policies, roles, risk tolerance, oversight	Zero-trust posture: device identity fingerprint in health and MQTT, Hermes peer mesh with HMAC and mTLS, authenticated agent-to-agent queries through the proxy. Privacy-by-design: no cloud API keys in the edge image; edge precomp runs first; cloud queries are signed. Storage governance via prune-first soak workflows. Audit trail via debug event ring and operator TAP artifacts.
Map	Context, asset categorization, risk identification	RF-DETR precomp for person and transit-safety scenarios. Optional nist_fr plugin applies NIST FATE presentation-attack detection (depth-based liveness) and quality scoring on depth-capable hardware. Justified H.265 recording with signed sidecar metadata. ONVIF Profiles S/T for interoperable video and event integration.
Measure	Metrics, testing, continuous evaluation	Soak gates and vision-verify scripts; blur and sharpness thresholds from FATE quality assessment; default detection confidence 0.35 for person class. Unit tests for MCP contracts, device identity, spatial fusion, and fleet schema. Harness validation of precomp counts. Thermal, encode-failure, and storage gates during sustained runs.
Manage	Risk treatment, monitoring, response	Prune-first policy in orchestrated soak batches; recovery scenarios after encoder stalls. Hermes self-heal and peer status. Cloud escalation (signed object-storage upload and generative or structured video APIs) only on justified clips via the proxy. ONVIF Events surface person detections for third-party alerting. Precomp authority prevents bad counts from reaching manage decisions.

FATE and FRTE in video surveillance

The optional nist_fr plugin targets Shadow hardware with native depth (for example Shadow 2). It implements controls drawn from NIST Face Analysis Technology Evaluation (FATE):

• Presentation-attack detection (PAD): Depth variance on face crops; low variance indicates a likely spoof (photo or screen).
• Quality assessment: Laplacian sharpness scoring; blurry or low-quality crops are rejected before publish.

Face crops publish to MQTT only when liveness passes and quality meets the configured minimum. Face Recognition Technology Evaluation (FRTE) applicability: these gates harden face-recognition evaluation in surveillance feeds by pairing RF-DETR person context (class 0) with liveness- and quality-filtered face events.

Video-surveillance cyber practices covered here:

• Authoritative on-device person counts before any cloud call
• Justified, signed recording metadata with device fingerprint
• Standard ONVIF surfaces instead of vendor lock-in
• Zero-trust transport (mTLS, HMAC, fingerprinted identity)
• Edge-only precomp; proxy-owned signed cloud escalation
• Blur and quality gates blocking low-confidence clips

RF-DETR model card hook

See Model card — RF-DETR for the full card. Key compliance hooks:

Parameter	Value	Role
Person class ID	0	RF-DETR label for person; matches training labels
Default confidence	0.35	Minimum score for person counting and event publish
Detection stride	~10 Hz	Independent substream on device DSP
Precomp authority	Always first	Injected into agent fusion, MCP tools, MQTT, and ONVIF Events; LLM layers do not override counts

Intended use: On-device person detection and optional weapon classes for sovereign video surveillance and mass-transit safety. Limitations: Depth-based FATE requires depth-capable hardware; host-side tracking may substitute when device object tracking is unavailable.

Detection and compliance flow

Next steps

• Model card — RF-DETR — class IDs, confidence, limitations
• ONVIF reference — person events from precomp detections
• Gate interpretation — soak PASS/FAIL semantics
• External references — OWASP, ONVIF, and cloud video APIs

Operator depth

Live fleet state, harness evidence, and gate chronology live in private operator handoff (not published).