PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantifying the Privacy Posture of Operator-Side 5G/O-RAN Profiles

Nikolaos Kekatos, Apostolos Valiakos, Alexios Lekidis, Elpiniki Papageorgiou

arXiv:2610.07976v1cs.CRcs.NI

TL;DR

Operator-side 5G/O-RAN profiles contain personal data, so the paper asks how much re-identification risk remains after standard syntactic anonymisation. It quantifies privacy posture and evaluates transformations against disclosure bounds and simulated adversaries, finding that generalisation matters more than pseudonymisation alone and that the mean-class-size PPI tracks the evaluated measures more closely than the kmin-based index.

  • Problem

    Operator-side 5G/O-RAN profiles can contain personal data, and the paper examines residual re-identification risk after standard syntactic anonymisation.

    Source: §I · p. 1

  • Method

    The paper computes k-anonymity, ℓ-diversity, t-closeness and PPI, then evaluates transformations using an exact-QI bound and two simulated adversaries.

    Source: §I · p. 1

  • Results

    Pseudonymisation alone leaves re-identification unchanged, while mean-class-size PPI′ is strongly associated with all three disclosure and attack measures (r=0.63–0.79).

    Source: §I · p. 2

  • Takeaways & Limitations

    The kmin-based PPI is not a security bound for the partial-knowledge attack, while PPI′ tracks the evaluated disclosure and attack measures more closely.

    Source: §I · p. 2

    The evaluation is modest in scale and the downstream utility result uses a single fixed stratified split, so trends are indicative rather than statistically powered.

    Source: H. Discussion and Limitations · p. 8

Abstract

from arXiv · show

Operator-side network profiles derived from 5G/ORAN traffic carry personal data such as ephemeral subscriber identifiers, slice-level KPIs, and control-plane signalling, and must be anonymised before release to a federated-learning aggregator, threat-intelligence exchange, or ML training pipeline. We study how much re-identification risk remains after standard operator-side anonymisation. We quantify privacy posture with k-anonymity, l-diversity and t-closeness, aggregate them into a composite Privacy-Posture Index (PPI), and measure residual re-identification across eight transformation configurations on internal PCAP captures and the public Idaho Labs 5GAD corpus, under a full-QI syntactic bound and two simulated adversaries. The evaluation is modest in scale, and we read its trends as indicative rather than definitive. Three findings emerge. Pseudonymisation alone leaves re-identification unchanged; material privacy gains arise when quasi-identifiers are coarsened through generalisation, optionally combined with suppression. A downstream classification task then shows that suppression-heavy releases retain majority-class utility but sacrifice much of their minority-class recall, a cost the aggregate metrics hide. Finally, the standard kmin-based PPI correlates only modestly with the disclosure bound and not at all with the partial-knowledge attack, whereas a mean-class-size variant PPI correlates strongly with all three disclosure/attack measures; we therefore read PPI as a regulator-facing summary, not a security bound. The profiles are produced by passive operator-side monitoring with rule-based DPI; our contribution is the privacy-quantification layer that computes these metrics, applies the transformation policy, and exposes both through an inspectable dashboard.

I. INTRODUCTION

5G/O-RAN profiles must support network operations while limiting exposure of personal data before release. The paper adds an operator-side privacy-quantification layer and evaluates how anonymisation changes re-identification risk and downstream utility.

  • The study quantifies profiles using k-anonymity, ℓ-diversity, t-closeness and PPI, assessing residual disclosure with an exact-QI bound and two simulated adversaries.PENM supplies the profiles; the paper's focus is the privacy-quantification layer operating on its output.Source: §I · p. 1
  • The evaluation covers 18 PCAPs—10 internal and 8 external Idaho Labs 5GAD Normal-1UE captures—across eight transformation configurations.Source: §I · pp. 1–2
  • Suppression-heavy releases retain majority-class accuracy but lose about half of minority-class recall; an exploratory label-aware policy recovers recall in the fixed split evaluated.The aggregate privacy metrics do not show this downstream utility cost.Source: §I · p. 2
  • Pseudonymisation alone leaves re-identification unchanged, while quasi-identifier generalisation, optionally with suppression, raises mean k appreciably above 1.The reported gains use IP-prefix generalisation with threshold suppression (C4) or protocol-count binning (C6).Source: §I · p. 2
  • The kmin-based PPI correlates only modestly with the disclosure bound, whereas mean-class-size PPI′ correlates strongly with all three measures (r=0.63–0.79).Source: §I · p. 2

II. BACKGROUND AND STATE OF THE ART

The background presents k-anonymity, ℓ-diversity and t-closeness as transparent syntactic privacy baselines, while noting their limits. It also situates PENM pseudonymisation choices across primitive, determinism and reversibility.

A. Privacy quantification: syntactic models

  • k-anonymity bounds identity-disclosure probability by 1/k on selected quasi-identifiers, ℓ-diversity requires sensitive-value diversity within classes, and t-closeness bounds class-to-global distribution distance.The paper treats these models as a transparent baseline alongside cryptographic pseudonymisation; differential privacy is left to offline post-processing.Source: A. Privacy quantification: syntactic… · p. 2

B. Cryptographic pseudonymisation

  • PENM's pseudonymisation choices span primitive, determinism and reversibility/trust distribution; deterministic HMAC supports longitudinal flow analysis, while IPv4-shaped pseudonyms preserve dashboard-compatible formatting.Source: B. Cryptographic pseudonymisation · p. 2

C. Privacy-aware network profiling

The paper addresses the gap in inspectable operator-side privacy quantification for 5G/O-RAN profiles. Its workflow transforms profile fields and reports privacy metrics before release.

C. Privacy-aware network profiling

III. NETWORK PROFILES

  • PENM derives one per-host record from passive packet capture using a port- and DPI-based role heuristic, including protocols, traffic counts and UE identifiers.The privacy layer consumes these structured records rather than the underlying PCAP.Source: §III · p. 2

A. Workflow

  • The workflow ingests a PCAP or tabular dataset, classifies fields as identifiers, quasi-identifiers or sensitive attributes, applies transformations, then reports before/after metrics and suggestions.Source: A. Workflow · p. 2

B. Privacy metrics

  • For each release, PENM partitions records by quasi-identifier tuple and computes k, ℓ and t; t uses total-variation distance between sensitive-attribute distributions.Rule-based suggestions include subnet aggregation for raw IP quasi-identifiers and threshold suppression when t exceeds 0.5.Source: B. Privacy metrics · pp. 2–3

C. Transformation primitives

The transformation layer offers pseudonymisation, generalisation and suppression. The PPI combines k, ℓ and t into a regulator-facing score, with a mean-class-size variant intended to reflect more of the class-size distribution.

C. Transformation primitives

  • PENM supports keyed HMAC, salted hashing, random tokenisation and IPv4-shaped HMAC pseudonymisation, with deterministic, per-pilot or per-run scope policies.The prototype uses 48-bit HMAC tokens for dashboard compactness; production should use 96–128 bits.Source: C. Transformation primitives · p. 3
  • Generalisation profiles include IP-prefix reduction, numeric binning, text-prefix masking and protocol-family abstraction; suppression can replace values directly or hide values below a support threshold.Threshold-based suppression is intended to prevent rare values from pinning k at 1.Source: C. Transformation primitives · p. 3

D. Privacy-Posture Index (PPI)

  • PPI combines capped k- and ℓ-based scores with 1−t using weights summing to one; its illustrative targets and default weights are conventional, not regulatory thresholds.PPI′ substitutes record-weighted mean class size for kmin because kmin ignores the broader class-size distribution seen by averaged attacks.Source: D. Privacy-Posture Index (PPI) · p. 3

E. Operator interface

The dashboard exposes transformation policies and their privacy effects for operator inspection before release.

E. Operator interface

  • Operators select identifiers, quasi-identifiers, and the sensitive attribute, then configure the transformation policy.The interface also provides rule-based suggestions.Source: E. Operator interface · p. 3
  • Before-and-after k, ℓ, and t metrics appear alongside a preview of transformed records.The preview is exportable.Source: E. Operator interface · p. 3
  • Each transformation and its metric effects remain visible rather than hidden behind a single anonymisation action.Source: E. Operator interface · p. 3

V. EXPERIMENTAL EVALUATION

The evaluation frames release risk under a trusted operator and an untrusted recipient, then tests three datasets. In the 12-session O-RAN example, the release has k=3, ℓ=2, and t=0.25, but the authors treat it as illustrative rather than statistically conclusive.

A. Threat model

  • The threat model assumes a trusted operator and an untrusted recipient who knows the transformation policy and selected attributes; it evaluates a full-QI bound and two simulated adversaries.The full-QI bound is 1/|E|; the partial-QI adversary knows each QI value with probability 1−p, using p=0.5.Source: A. Threat model · p. 4

B. Setup and datasets

  • The evaluation uses an O-RAN testbed, an EV-charging dataset, and ten internal PCAPs, with QIs and sensitive attributes varying by configuration.The O-RAN scenario contains 12 sessions; the EV-charging dataset contains 513 sessions.Source: B. Setup and datasets · p. 4

C. O-RAN testbed results

  • The O-RAN release has k=3, ℓ=2, and t=0.25, with its smallest class containing only two sensitive-attribute values.The authors call this 12-session, three-class example illustrative, not a basis for statistically significant conclusions.Source: C. O-RAN testbed results · p. 4

D. EV-charging results

The EV-charging transformation improves the reported privacy metrics while retaining about 14% of rows. Across the PCAP configuration sweep, generalisation and suppression reduce disclosure, while the kmin-based PPI aligns less well with simulated attacks than its mean-class-size variant.

D. EV-charging results

  • k rises from 1 to 5, ℓ from 1 to 2, and t falls from 0.90 to 0.73 as the EV-charging release shrinks from 513 to 70 rows.The policy combines HMAC pseudonymisation, generalisation, and suppression of equivalence classes smaller than five; about 14% of rows remain.Source: D. EV-charging results · p. 4

E. Adversary-success simulation and PPI validation

  • The full-QI bound assumes knowledge of every QI value, while the partial-QI and linkage attacks use simulated incomplete or probabilistic attribute information.Both simulated attacks average over 20 fixed-seed Monte Carlo repetitions, with every released row as a target.Source: E. Adversary-success simulation and… · p. 5
  • HMAC pseudonymisation alone leaves QI equivalence classes and privacy measures unchanged, while adding threshold suppression to /24 generalisation raises mean per-release kmin to 3 and cuts exact-QI re-identification by ∼72 percentage points versus baseline.C6 has the lowest residual re-identification: 25.3% exact and 22.3% partial.Source: E. Adversary-success simulation and… · p. 5
  • The exact-QI disclosure probability averaged over records equals the equivalence-class granularity ratio |E|/N, so the paper does not treat that ratio as an independent utility metric.Task-level utility is measured separately with the downstream classifier.Source: E. Adversary-success simulation and… · p. 5
  • PPI′ correlates with all three disclosure or attack measures (combined r=+0.79 exact, +0.63 partial, +0.75 linkage), while PPI correlates only modestly with the exact-QI bound (r=+0.30) and is essentially uncorrelated with the partial-QI attack (r=+0.08).The paper retains PPI as a kmin-oriented dashboard indicator and presents PPI′ as an attack-alignment variant, not a replacement regulatory metric.Source: E. Adversary-success simulation and… · p. 5, Figure 2 · p. 6
  • C4 transformation processes 10^2, 10^3, and 10^4 synthetic records in 0.3, 3.2, and 32 ms, respectively.At the evaluated record volumes, the transformation stage is well below the per-PCAP DPI/profiling cost.Source: E. Adversary-success simulation and… · p. 5

F. External validation on 5GAD

On the independent 5GAD corpus, the configuration ranking matched the internal corpus, while exact-QI re-identification was lower for C4 and C6. The authors relate this difference to larger equivalence classes in 5GAD and note that kmin alone can misrank releases.

F. External validation on 5GAD

  • The external sweep used eight independent 5GAD Normal-1UE profiles, with 5–322 records each, compared against ten internal PCAPs on the same configurations.Larger captures were limited to 200,000-packet slices; the external testbed had no shared topology, addressing plan, or role-classifier ground truth with the internal one.Source: F. External validation on 5GAD · p. 6
  • C4 and C6 outperformed other configurations on both corpora, while C0 and pseudonymisation-only C1 coincided on every metric.HMAC pseudonyms preserve the equivalence-class partition by construction.Source: F. External validation on 5GAD · p. 6
  • 5GAD exact-QI re-identification was 5.0% for C6 and 11.4% for C4, versus 25.3% and 28.0% internally, respectively.The authors attribute the gap to richer protocol-count and protocol-set distributions and larger QI/SA partitions in 5GAD.Source: F. External validation on 5GAD · p. 6, Table IV · p. 7
  • 5GAD had no larger mean kmin than the internal corpus despite lower exact-QI re-identification, because kmin misses the larger classes containing most records.The authors identify kmin as the smallest class size and krw as the record-weighted mean class size; they associate expected exact-QI success with krw rather than kmin.Source: F. External validation on 5GAD · p. 6
  • Across 144 releases, PPI correlated with 1−exactReID at r=+0.30, below the internal-only r=+0.45; the mean-class-size PPI′ closes this gap.PPI’s k-component rewards kmin, whereas exact-QI disclosure and simulated attacks depend on the broader class-size distribution.Source: F. External validation on 5GAD · p. 6

G. Downstream ML utility under suppression

Suppression-heavy releases preserved overall classification accuracy but disproportionately removed minority-role examples, sharply reducing their recall. An exploratory label-aware policy recovered rare-role utility in the evaluated fixed split, but trades privacy on rare records.

G. Downstream ML utility under suppression

  • The evaluation pooled 664 per-IP records, trained a random forest on survivors after QI-support suppression, and tested on the untouched stratified 30% split.The roles rarer than six records were grouped into Other; the task predicted node role from behavioural features.Source: G. Downstream ML utility under… · p. 7
  • At τ=8 with 19% survival, overall accuracy stayed near 0.76, while macro-F1 fell to 0.37 of baseline and rare-role recall dropped from 0.92 to 0.42.Beyond about 15% survival at τ=12, the model collapsed; suppression preferentially removed sparse equivalence classes where minority roles lived.Source: G. Downstream ML utility under… · p. 7
  • Exempting the two rarest role labels from suppression restored edge-case recall from 0.42 to 0.92 and macro-F1 retention from 0.37 to 0.68 at τ=8.Survival rose from 19% to 37%; the authors describe this fixed-split, label-aware policy as an exploratory upper bound that trades privacy on rare records for utility.Source: G. Downstream ML utility under… · p. 7

H. Discussion and Limitations

The privacy layer provides syntactic metrics but no formal differential-privacy guarantee, and evaluates releases individually rather than accounting for composition across repeated publications. The study’s scale, traffic coverage, and downstream evaluation also constrain how broadly its results can be applied.

H. Discussion and Limitations

  • The privacy layer is purely syntactic and does not offer a formal differential-privacy guarantee.The authors also note critiques of syntactic anonymisation,.Source: H. Discussion and Limitations · p. 7
  • Assessing each release in isolation leaves privacy degradation through composition across sequential re-publications unaccounted for.The paper identifies repeated publication of the same profiles as a setting where composition can matter.Source: H. Discussion and Limitations · p. 7
  • The evaluation is modest in scale, and the O-RAN snapshot and benign-only 5GAD corpus leave attack-labelled, multi-UE, and partially encrypted traffic unevaluated.The authors characterize results as indicative rather than statistically powered and scope a larger multi-capture, multi-UE evaluation as future work.Source: H. Discussion and Limitations · pp. 7–8
  • The downstream results use one fixed stratified split, while the utility-aware suppression result is label-aware and an exploratory upper bound.Repeated resampling is future work; profiler throughput is outside the reported transformation-latency measurement.Source: H. Discussion and Limitations · p. 8

VI. CONCLUSIONS AND FUTURE WORK

The paper presents an operator-side privacy-quantification layer for 5G/O-RAN profiles, combining syntactic metrics and transformations in an inspectable dashboard. Its evaluation includes an O-RAN testbed and an EV-charging dataset, while further privacy-model and deployment extensions remain open.

VI. CONCLUSIONS AND FUTURE WORK

  • The contribution is a privacy-quantification layer combining k/ℓ/t metrics into PPI with pseudonymisation, generalisation, and suppression policies in an inspectable dashboard.The conclusion distinguishes this layer from the network-profile production itself.Source: §VI · p. 8
  • In the EV-charging evaluation, transformation raised k from 1 to 5 and ℓ from 1 to 2, and reduced t from 0.90 to 0.73 at about 86% record suppression.The evaluated settings also included an O-RAN testbed scenario with k=3, ℓ=2, and t=0.25.Source: §VI · p. 8
  • Proposed extensions include partition-based k-anonymisation, newer syntactic models, per-record risk estimators, and formal bridges between syntactic privacy measures and differential privacy.The authors also discuss applying differential privacy, federated learning, and cryptographic audit methods offline.Source: §VI · p. 8
Loading 2610.07976v1…