| Internet-Draft | Semantic Sensing Privacy | November 2026 |
| Karahan | Expires 20 May 2027 | [Page] |
Radio sensing in ISAC and AI-assisted networks produces semantic outputs such as target position, velocity, class, and identifiers. These outputs can reveal sensitive information about people who take no part in any communication service, especially when they are released repeatedly, combined across sources, or shared across administrative domains. This document describes the privacy risks of releasing such outputs and the network management functions needed to control their release: policy, enforcement at trust boundaries, consumer authorisation, accounting, and auditing. It surveys protection approaches and lists open research questions. It does not define a protocol or an algorithm.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 20 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Networks can now infer the state of physical objects from radio observations. Integrated sensing and communication (ISAC), joint communication and sensing, and AI-assisted radio sensing can provide target detection, position, velocity, motion class, and uncertainty estimates. Applications range from cooperative transportation to industrial automation and digital representations of the physical environment.¶
Sensing outputs differ from ordinary application data in one important way: the network may describe a person who supplied no data, has no relationship with the requesting application, and does not know that sensing takes place. A position estimate shows presence at a place. Repeated position and velocity reports can reveal trajectories and visits. Even without a name, a stable identifier, precise timestamps, or motion patterns can be enough to link reports and re-identify a person.¶
Privacy problems can arise at every stage of the data lifecycle. Radio observations may be retained. Sensing features may be combined with other sources. AI models may derive higher-level attributes. Applications may ask for results at different levels of detail. Encrypting the transport does not decide whether an authorised recipient should get precise target information, and perturbing one position field does not protect a record whose other fields are correlated with it.¶
Existing guidance such as [RFC6973] gives general principles for privacy analysis. NMRG work on AI and network management lists data privacy, heterogeneous telemetry, distributed AI, and operational integration as research concerns [I-D.irtf-nmrg-ai-challenges], and the distributed AI deployment work shows why it matters where data are processed and what crosses administrative boundaries [I-D.irtf-nmrg-ai-deploy]. Semantic sensing is a concrete case where these concerns meet.¶
This document therefore looks at how the disclosure of derived sensing information is managed. It does not propose a new sensing waveform or location-privacy algorithm. Its intended contribution to NMRG is the management of privacy controls across network functions, consumers, and administrative domains.¶
Section 2 gives objectives, scope, and terminology. Section 3 describes the sensing pipeline and its privacy threats. Section 4 surveys protection approaches. Section 5 covers architecture and management functions. Section 6 relates the topic to NMRG work. Section 7 lists open research challenges, Section 8 gives illustrative use cases, and Sections 9 and 10 cover security and IANA considerations.¶
This document aims to:¶
The scope starts with sensing observations and derived records and ends with the decision to disclose a result to a consumer. The focus is on outputs that describe individual targets or small groups in fine detail. Both network-internal consumers and external applications are considered.¶
This document does not define a protocol, message schema, interface, management information model, cryptographic primitive, or radio measurement procedure. It does not require any particular network function to implement a privacy enforcement point. It does not provide a legal basis for processing personal data, describe how to obtain consent, or recommend a universal privacy parameter. The use cases are illustrative and are not evidence of implemented systems or measured performance.¶
A sensing system turns raw radio observations into features, and features into an inferred physical state. Examples are channel observations, range-Doppler features, detection results, coordinates, and trajectories. AI components can combine measurements over time or across sources to add semantic detail.¶
Figure 1 shows a logical pipeline. It is not a protocol stack and does not imply that each function runs on a separate node.¶
Radio observations
|
v
Detection, tracking, AI processing
|
v
Semantic target records
|
v
+--------------------------+ +----------------------------+
| Privacy Enforcement |<----| Policy, consumer |
| Point (logical PEP) | | authorisation, audit |
+------------+-------------+ +----------------------------+
|
| permitted output
v
Consumers (applications, network management functions)
A PEP acts when outputs are delivered, but earlier controls also matter. Limiting raw-data retention and avoiding unnecessary tracking reduce exposure before any release decision is made. Where the disclosure boundary lies depends on where the producer, the trusted domain, and the consumers are deployed.¶
Position shows where a target was observed. Velocity and heading show motion and, with timestamps, constrain past and future positions. Target class shows whether the target is a pedestrian, a vehicle, or something else. Confidence fields show how far an observer can rely on a report. Persistent identifiers make reports easy to associate. Fields that are harmless alone can still allow re-identification or precise tracking when combined.¶
Sensitivity depends on context. The same spatial precision may be harmless for an industrial robot in a test hall and sensitive for a pedestrian near a home. Policy therefore has to consider the subject, the purpose, the attributes released, and the disclosure context, not only spatial accuracy.¶
The consumers considered here are authorised but honest-but-curious: they follow the rules but try to infer more than their service needs. Unauthorised recipients that obtain data through a failure of access control are also considered. A consumer may hold maps, movement models, records from other sources, or earlier releases, and may combine overlapping observations received through different applications.¶
The relevant attacks are:¶
A PEP can follow its policy correctly while a consumer still infers sensitive information from permitted releases. Correct policy execution and meaningful privacy protection are different properties. If a producer or PEP is compromised, output-level protection is bypassed entirely, and security measures outside any statistical release mechanism are needed (Section 9).¶
A mechanism that works for one report can fail over many. Averaging independently perturbed measurements of a stationary target reduces the estimation error. For a moving target, temporal filtering and physical motion constraints make a series of noisy positions more informative than any single one. Linking attacks do not require a persistent identifier, although an identifier makes them easier.¶
Privacy analysis therefore needs to state the release window, the linkability assumptions, and the side information available to the consumer. Event-level guarantees and trajectory-level goals are different and must not be used interchangeably. This matters most where continuous sensing is offered as a managed service.¶
No single protection covers every attribute and threat. The approaches below are complementary. They can be selected and combined according to the consumer's legitimate need, the trust boundaries, and the privacy and utility requirements.¶
An application often needs fewer fields than a sensing function produces. A traffic-density service may need regional counts rather than coordinates and identifiers. A policy can limit the fields released, coarsen the location, lower the update rate, omit stable identifiers, or suppress small groups whose members could be singled out.¶
Aggregates such as occupancy, density, flow direction, and event counts reduce direct exposure of individual records. A purpose-limited view restricts what a consumer may request and how often.¶
These measures are not formal guarantees. A coarse location combined with exact velocity and frequent timestamps can still allow tracking. Small populations, differencing of overlapping queries, and repeated snapshots allow inference from aggregates. Suppression can itself reveal that a target is present. Thresholds and update rules should be evaluated against realistic adversary access, and views for different consumer classes must be enforced consistently and be auditable.¶
Rotating or omitting identifiers makes explicit linkage harder, but a trajectory can still be linked through physical continuity, timestamps, and location. Expiry of records, retention limits, access logging, and limits on onward disclosure address other dimensions of risk. They remain useful when a formal perturbation mechanism is also used.¶
Differential privacy (DP) bounds how much a mechanism's output distribution can change between adjacent inputs [DWORK06]. The definition of adjacency decides what is protected: add/remove-record adjacency, substitution adjacency, and metric proximity protect different things.¶
Geo-indistinguishability adapts a DP-style guarantee to a location metric [GEOIND]. For a randomised mechanism M, locations x and x', an output set S, and a distance d, it requires¶
Pr[M(x) in S] <= exp(epsilon * d(x,x')) * Pr[M(x') in S].¶
A smaller epsilon gives stronger indistinguishability at a fixed distance and generally more perturbation. Epsilon has units of inverse distance when d is measured in distance units. The property constrains output distributions. It does not make estimating a location impossible.¶
For three-dimensional Euclidean coordinates, an isotropic mechanism has a noise-vector density proportional to exp(-epsilon * ||z||_2). The direction of the noise is uniform on the sphere, and its radius follows a gamma distribution with shape 3 and rate epsilon, so the mean radius is 3/epsilon. The three coordinate components are not independent one-dimensional Laplace samples. Earlier work on three-dimensional geo-indistinguishability is in [YAN2024].¶
Choosing epsilon only from a tolerated mean error is not a sufficient policy decision. Setting epsilon = 3/A gives a mean perturbation radius of A. It does not bound the error and does not make a given inference attack weak. The calibration must also consider the full error distribution, the baseline sensing error, the adversary's side information, and the application's requirements. Unbounded noise cannot promise that every perturbation stays below a strict safety threshold.¶
The guarantee covers only the chosen location mechanism on its input and output space. Detection state, exact velocity, stable identifiers, repeated reports, and unprotected correlated features are not covered.¶
Approaches can be combined, for example by aggregating a target set, suppressing identifiers, lowering the update rate, and perturbing coordinates where individual locations must still be released. Such a chain does not automatically give an end-to-end guarantee. The adversary's access to original and derived outputs must be specified first. Mechanisms should be compared on the protected subject or event, the formal guarantee or policy goal, utility, computational cost, update rate, and operational constraints.¶
A privacy management function could offer an approved set of release profiles instead of exposing raw parameters to consumers. A profile might fix the allowed fields, spatial granularity, maximum sampling rate, retention policy, and accounting method. A common representation of such profiles is a research question, not something defined here.¶
A PEP is not necessarily a separate device or protocol endpoint. It is a logical function at the boundary where data leave a trusted domain. It can check authorisation and purpose, build the permitted view, apply an approved transformation, record the release, and deliver or reject the request. It cannot protect against information that the same consumer obtains through another route.¶
The architecture should say who defines policy, who evaluates it, who applies transformations, and who audits decisions. A small deployment may combine these roles; a large one may separate them. A PEP late in the pipeline protects downstream consumers but not exposure inside the pipeline, so trust assumptions and data-flow visibility should be explicit.¶
In a centralised arrangement, precise outputs go to a trusted service that applies privacy controls before publishing. This simplifies consistent enforcement but enlarges the trusted collection domain and may add communication cost and latency. In an edge-based arrangement, filtering or perturbation happens near the producer and only approved outputs travel onward. This reduces the movement of precise data but requires consistent distribution of policy to many enforcement points. A cross-domain arrangement adds a second administrative domain with its own policy.¶
Figure 2 shows the three arrangements. Arrows show possible disclosure paths, not a required protocol or a 3GPP functional mapping.¶
(a) Centralised (b) Edge-based
Sensing sites Site A Site B
| | |
v v v
Trusted collector PEP A PEP B
| | |
v +------+-------+
Central PEP |
| v
v Consumers
Consumers
(c) Cross-domain
Domain A Domain B
+------------------+ +------------------+
| Sensing + PEP A |------------->| Policy gateway / |
| Local policy A | protected | release controls |
+------------------+ output | Local policy B |
+--------+---------+
|
v
Consumers B
Deployments can mix these arrangements. The key question is whether sensitive or precise outputs cross a boundary before the relevant policy is applied, and whether a consumer can tell what the received data mean and what their limits are.¶
Access control has to address more than whether a consumer may receive any data. The same consumer may be allowed regional occupancy but not individual trajectories. The authorisation context can include identity, declared purpose, role, area, period, output attributes, precision, and update rate. Limits on onward sharing and retention should be part of it where they can be enforced.¶
Consumer classes need different output products. A traffic planner can usually work with aggregate flows, while an operational controller may need timely state information. Such differences are not blanket permission to release exact personal data. Access needs and lawful authority must be assessed separately.¶
An operator may run many PEP instances for different sensing sites, edge clusters, or applications. Policies must be distributed, versioned, checked for conflicts, and updated when authorisations or service requirements change. A PEP needs a defined response to missing, expired, or inconsistent policy, such as withholding detail, returning a lower-precision view, or escalating. This document does not specify one fallback for all cases.¶
Updates to models or sensing functions can change the accuracy or content of outputs and so require reassessment of release profiles. A management workflow should tie model lifecycle changes to policy validation and monitoring. This matches the NMRG concern with integrating AI into existing management procedures [I-D.irtf-nmrg-ai-challenges].¶
One consumer may receive a sequence of reports, and several consumers may request overlapping views of the same target. If noise is applied independently per request and the reports are collated, more is revealed than intended. Accounting should distinguish independent observations, repeated releases of the same information, and releases to separate but possibly colluding recipients.¶
A formal privacy budget has meaning only for a specified mechanism, adjacency relation, and composition model. Rate limits and caps on the number of authorised views are policy limits, not formal DP budgets. The management question is how an operator can assess cumulative disclosure without keeping sensitive identifiers longer than needed, and without letting the accounting system become a tracking tool.¶
Consumers may need metadata on units, coordinate reference system, timestamp precision, quality indicators, permitted use, and whether the released state has been altered. Revealing that a mechanism is used, and its settings, does not by itself weaken a properly defined DP guarantee, but excess metadata can reveal protected information through correlation. Metadata minimisation and quality-of-information reporting need to be assessed together.¶
Observability can record the policy version, the consumer, the output category, the enforcement status, and aggregate counters. Audit logs should not by default contain unperturbed target records. Retention periods, log granularity, access restrictions, and deletion procedures are deployment decisions. An audit trail shows that a rule was applied. It does not show that the rule gives enough privacy.¶
Domains differ in trust, purpose, retention rules, and protection methods. One domain may supply coarse occupancy while another supplies target-level records, and combining them can defeat the assumptions under which each was released. A domain boundary therefore calls for a new policy evaluation and for a description of the assumptions that travel with a dataset.¶
Research is needed on finding compatible release policies and on accounting for disclosures across domains without assuming one global trusted operator. Intent and policy mechanisms in network management [RFC9315] are a possible conceptual basis. This document does not map to any management protocol.¶
Sensing is often requested to support actions, not just to publish data. More uncertainty or a lower reporting rate can degrade target association, collision warning, equipment monitoring, or control quality. Policy should separate safety-critical real-time consumers from analysis applications that can accept coarse or delayed data. Being safety-critical does not by itself entitle a consumer to unrestricted data.¶
A privacy-aware service may have to represent both sensing uncertainty and privacy-induced uncertainty. These do not simply add as scalars; covariance, nonlinear tracking, and the downstream estimator may matter. Where hard accuracy or timing requirements exist, independent unbounded perturbation may be unsuitable without separate error handling, risk assessment, or a different privacy strategy.¶
The NMRG study of AI and network management identifies challenges in data representation and privacy, distributed computation, operational integration, trust in AI outputs, and automation [I-D.irtf-nmrg-ai-challenges]. The distributed AI deployment draft examines how placement across local, edge, and cloud resources affects how information is handled [I-D.irtf-nmrg-ai-deploy]. Semantic sensing privacy adds a related management problem: how to govern the release of AI-derived information about physical targets across those environments. Neither document addresses radio sensing or ISAC specifically. This document does not extend their scope; it takes semantic sensing outputs as one concrete case of the privacy and data-handling concerns they raise.¶
These documents have the following points of contact with this one:¶
None of this implies that the existing NMRG documents already solve semantic sensing privacy. A useful research outcome would be a common description of the disclosure problem: what must be protected, what a consumer legitimately needs, which controls are applied, and what remains exposed.¶
A formal privacy statement must say whether it protects target presence, a single position, a whole trajectory, a person's participation, or a group. Each goal needs its own definition of neighbouring inputs and probably its own mechanism. The definitions must allow for sensing errors, missed detections, and targets appearing and disappearing.¶
Position and velocity are linked by physical dynamics. Perturbing position while sending exact velocity weakens location protection. Perturbing both independently can create inconsistent or physically impossible trajectories. Joint release methods are needed with explicit guarantees and explicit utility.¶
An event-level mechanism does not give trajectory-level privacy. Open questions include how to allocate disclosure across sampling times, how to handle long-lived targets, and whether temporal structure can be preserved without allowing averaging or filtering attacks. Prior information and tracking dynamics should be part of the evaluation.¶
Rotating identifiers limits direct linkage but cannot remove association through motion and context. A deployment may receive outputs from overlapping radio sensors, cameras, and network management applications. Cross-source correlation and authorised data fusion belong in the privacy analysis as normal cases.¶
A shared vocabulary for consumers, purposes, protected attributes, precision, frequency, retention, and transformation profiles could let heterogeneous systems interpret policy consistently. It is not clear which parts can be expressed independently of the sensing technology and which need domain-specific detail. This document proposes no wire encoding.¶
Several domains may apply different controls to overlapping data. Open questions include policy conflicts, accounting for combined releases, auditing without excessive retention, and which constraints can be checked locally and which need coordination.¶
Applications need enough information to judge output quality. The uncertainty of a private location includes sensing error and protection-induced error, but detailed metadata or repeated confidence updates can leak information. Research should determine which uncertainty and quality indicators can be shared safely and usefully.¶
Mechanisms should be compared under realistic reporting rates, target densities, sensing uncertainty, mobility, map constraints, and colluding consumers. Useful metrics include re-identification and tracking success, quality of aggregate statistics, localisation error distributions, application task performance, computational overhead, and enforcement failures. Simulations and synthetic trajectories support reproducibility. Public release of real sensing records needs its own privacy assessment.¶
Road-side units detect moving objects and derive position, direction, and motion state. A road-operations application may need only counts or flows, while a cooperative safety application needs timely detail. The management problem is to give each consumer a view that fits its purpose, without routinely exposing precise, linkable pedestrian or vehicle trajectories to all of them.¶
An edge PEP can publish aggregate flows for planning and a separate authorised state view for safety processing. The management system must specify which consumers get which fields, how often, and under what conditions. An evaluation should test whether removing identifiers and reducing precision really reduces linkage when external road maps and repeated reports are available.¶
Factories and campuses can use radio sensing for asset monitoring, occupancy, and safety awareness. Staff, visitors, and equipment share the space, and several administrative or application domains may be involved. A production-monitoring application may need machine occupancy or movement events but not identifiable worker trajectories.¶
Controls near the producer can release aggregated occupancy or event alerts to wider IT systems and keep target-level data in a restricted domain. The example shows the need for policy differentiation, retention rules, auditing, and attention to site safety. Technical perturbation does not settle employment or legal privacy obligations.¶
AI-assisted management can combine network telemetry with environmental observations to understand service conditions or simulate network responses. A management function may benefit from regional mobility trends or occupancy estimates without needing individual target histories. Network digital twins and planning tools may take in sensing-derived features alongside topology and performance data.¶
A privacy-aware interface can limit releases to what planning needs, while a separate trusted operational context handles any justified finer-grained state. AI models can infer sensitive properties from apparently coarse features, so evaluations must consider the combined input and accumulated history of the management application. This case most directly links semantic sensing privacy to NMRG's AI and network management work.¶
This document describes a privacy-management problem and introduces no protocol. Deployments of disclosure controls still need security properties that keep the assumptions of the privacy analysis valid.¶
These considerations should be read together with general Internet privacy guidance [RFC6973]. They do not replace jurisdiction-specific data protection requirements.¶
This document has no IANA actions.¶