ahancock516/oai-b210-plkg-dataset

OAI 5G B210 PLKG Dataset

Single-UE (B210) capture campaign for the channel-informed neural PHY key-generation / RF-fingerprinting project. This is a device-specific companion dataset to ahancock516/oai-5g-plkg-dataset — it isolates data from a single B210 UE so it can be combined later with equivalent captures from other UE hardware (additional single-UE datasets to follow) into a multi-device dataset for validating IMSI/RNTI labeling and RF fingerprinting across device types.

Deployment

• OAI 5G NR SA: X410 gNB + B210 UE
• Band n78, 51 PRB, 30 kHz SCS (numerology 1), DDSUU TDD pattern
• gNB: DGX Spark (ARM64) host; UE: separate x86_64 host, same LAN

Fixed-PRB allocation (both directions target 20 PRB / 240 subcarriers)

• UL (PUSCH): every scheduled grant — including Msg3 — is unconditionally forced to 20 PRB (FIXED_PUSCH_RB_SIZE in the gNB MAC scheduler). 100% of captured UL records measure rb_size == 20.
• DL (PDSCH): forced to 20 PRB only for grants actually carrying DTCH (user-plane) data, detected per-slot from RLC buffer status (sched_ctrl->rlc_status) immediately before scheduling. Control-plane signaling (registration, PDU Session Resource Setup, RRC Reconfiguration, Msg3/Msg4 RA procedure) is left at its natural, unconstrained size — an earlier unconditional version of this patch silently starved those larger control messages and broke PDU session establishment. 99.6% of captured DL records measure nb_rbs == 20; the remainder are natural-sized control-plane grants (mostly 5 PRB).

Traffic generation

Captures were generated with sustained bidirectional iperf3 traffic (--bidir, UDP, 5 Mbit/s each direction) between the UE and the core network's external data network host, rather than repeated attach/detach cycling — this produces continuous real DTCH data grants on both UL and DL simultaneously, instead of relying on registration-phase signaling alone.

Contents

pusch/pusch_dataset_b210_bidir_20prb.bin   # UL captures, format v5
pdsch/pdsch_dataset_b210_bidir_20prb.bin   # DL captures, format v4
paired/paired_b210_bidir_20prb.h5          # UL/DL reciprocity pairs

Binary formats (PUSC/PDSC magic, versioned header) are documented in the respective capture-plugin repos: nr_pusch_capture, nr_pdsch_capture. Both plugins write a wall-clock timestamp_ns (CLOCK_REALTIME) per record, used for pairing.

Pairing method

Built with build_paired_dataset.py (--pairing-method timestamp_primary --comparison-domain delay): finds each UL capture's nearest DL capture by real timestamp per IMSI (not frame/slot distance, which is unreliable across the 10.24s SFN wrap period), then compares full-bandwidth delay-domain |h(tau)| rather than frequency-domain H(f) — valid even when the UL/DL grants share no overlapping subcarriers, which fixed-size scheduling makes common.

Known caveats

• Yield: 70,498 pairs from 77,787 UL / 64,991 DL captures — the highest yield observed in this project to date, attributable to sustained bidirectional traffic keeping UL and DL activity continuously co-occurrent in time (vs. registration-only cycling, where pairs only arrive in bursts around attach events).
• Reciprocity correlation: delay-domain magnitude correlation on a 3,000-pair sample: median 0.679 (mean 0.683, std 0.098; p10=0.559, p90=0.813). This is a real, non-degenerate reciprocity signal, but below the best value previously recorded on this rig (median 0.84, from a session with specifically re-tuned antenna positioning/gain). This capture was not re-tuned before the session — treat 0.68 as this session's as-found RF condition, not a ceiling.
• Cross-host clock sync: both hosts are individually NTP-synced, but to different upstream sources — the gNB host to a public NTP pool (1-2ms RMS offset), the UE host to a local LAN NTP server (10µs RMS offset). Cross-host timestamp alignment is bounded by the gNB host's accuracy, not sub-millisecond, even though the pairing search window is ms-scale.
• K-factor / residual-entropy check not yet applied. High UL/DL correlation is necessary but not sufficient for good key-generation material — a channel dominated by one strong deterministic (e.g. line-of-sight) component can correlate well while still being low-entropy and potentially predictable to an eavesdropper. This check is still open project work, not yet run against this dataset.

Intended use

• UL split (pusch/): shared with a collaborator for RF fingerprinting model training (independent of the reciprocity/pairing work above).
• Full dataset: validating IMSI-to-RNTI labeling and reciprocity pipeline correctness on a single, known device, ahead of combining with other UE hardware into a multi-device dataset.