Privacy-Preserving Intent Fulfilment and Assurance for 6G RAN

Fulfilment and assurance without traffic inspection

Intent-based management lets an operator state a service objective and leaves translation, provisioning and assurance to the system. Existing fulfilment and assurance approaches assume full observability: traffic classification, deep packet inspection, per-flow tracking or vendor-internal telemetry. Those requirements clash with data-protection rules in multi-tenant networks, with vendor opacity in multi-vendor O-RAN, and with traffic that is now mostly encrypted.

The architecture in this paper restricts what the coordinator sees. It provisions from a declared intent category, one of K classes, and verifies fulfilment using only cell-level standardised performance-management counters at the O1 interface. It never reads payloads, per-user traces or node-internal state.

A leakage ceiling set by the architecture

Because the allocation is a function of the category alone, a data-processing inequality bounds what it reveals about traffic content by log₂ K bits for K equally likely categories. Intent-traffic unlinkability and node-opaque verification then hold by construction. Privacy comes from not collecting the data, not from protecting it after collection.

Detection and privacy pull the same way. If verification needs at least q cells per category among M cells, the ceiling becomes log₂ ⌊M/q⌋ bits, so the detection requirement fixes the leakage ceiling without a separate privacy parameter. When individual records add state-independent noise to the category aggregate, a verifier on the category mean is at least as powerful as any verifier on individual records.

The paper maps the architecture onto the 3GPP intent lifecycle and the O-RAN Non-RT RIC. On production counters from four operator networks, detection recall plateaus beyond moderate numbers of categories while the leakage ceiling keeps rising. Under category-wide replacement the category-mean score has higher AUC on three networks and comparable AUC on the fourth. Per-cell screening suits sparse faults.

What the bound covers

The bound covers the category and resource channel. It does not bound everything a coordinator could infer from the PM counters themselves. The adversary is an honest-but-curious coordinator that follows the protocol. Side channels below the cell level, auxiliary data correlated with traffic, and a coordinator that bypasses the architecture are outside the model.

The evaluation uses injection-based ground truth, because the operator datasets carry no fault labels. These experiments do not establish the theorem's observation premise. Within-period co-movement of cells limits a category mean to the evidence of 6 to 14 independent cells, and category-level verifiers need per-category calibration.

Place in the research programme

The formal properties used here come from Information Requirements for Service Allocation and Aggregate Verification, which bounds the welfare cost of category-level provisioning and shows when verification on the aggregate loses nothing. This paper develops the privacy consequences of that restriction and places it in 6G intent management, part of my work on autonomous network management.