ahancock516/oai-n77-pusch-51prb-fingerprint-dataset

5G NR PUSCH RF Fingerprinting — 6 Devices, 3 Same-Model Pairs (fixed 51 PRB)

Uplink PUSCH IQ captures from six real 5G devices, recorded gNB-side on a live OpenAirInterface 5G SA network. 600,000 records (100,000 per device), every one captured on a single identical radio configuration.

Every record is a fixed 51 PRB uplink allocation — the full carrier width (612 subcarriers, 18.36 MHz at 30 kHz SCS).

This is one half of a paired-bandwidth experiment

A companion dataset covers the same six physical devices on the same band, same frequencies, same TDD pattern, same gNB hardware, differing only in uplink allocation width:

dataset	allocation	subcarriers
oai-n77-pusch-20prb-fingerprint-dataset	fixed 20 PRB	240
this one	fixed 51 PRB	612

Because everything else is held constant, the pair isolates occupied bandwidth as an independent variable. Questions the pair makes answerable:

• Does more occupied bandwidth improve separability (more of each transmitter's frequency response observed per record), or dilute the narrowband impairments a 20 PRB view concentrates on?
• Do models trained at one width transfer to the other, or is a learned fingerprint allocation-specific?
• Do the same-batch pairs — the hardest comparisons — benefit more or less from extra bandwidth than the easy across-class ones?

51 PRB is also the practical ceiling for the B210 units over USB 3.0 (three-quarter sampling at 23.04 MSPS), so 20 and 51 PRB bracket the usable range for this device set rather than being arbitrary points.

The three same-model pairs

pair	devices	relationship
Quectel RM520N-GL	unit A / unit B	same model, same manufacturing batch (near-sequential IMEIs)
Google Pixel 7	UE4 / UE5	same model, different units
USRP B210	unit 1 / unit 2	same model, same manufacturing batch (adjacent serials)

Separating two units of the same model from the same batch is far harder, and far more meaningful, than separating a phone from an SDR. Every pair was verified as physically distinct hardware via a hardware identifier — see Device identity verification.

Devices

#	device	class	IMSI (label)	hardware ID	file
1	Quectel RM520N-GL unit A	embedded modem	001010000000007	IMEI 868371056697259	pusch_dataset_rm520n_imsi007_51prb_100000labeled.bin
2	Quectel RM520N-GL unit B	embedded modem	001010000000006	IMEI 868371056697580	pusch_dataset_rm520n_imsi006_51prb_100000labeled.bin
3	Google Pixel 7 (UE4)	smartphone baseband	001010000000010	IMEISV 3536449372216689	pusch_dataset_pixel7_ue4_imsi010_51prb_100000labeled.bin
4	Google Pixel 7 (UE5)	smartphone baseband	001010000000008	IMEISV 3536449387387628	pusch_dataset_pixel7_ue5_imsi008_51prb_100000labeled.bin
5	USRP B210 unit 1	SDR (OAI UE stack)	001010000000001	serial 3113F40	pusch_dataset_b210_unit1_imsi001_51prb_100000labeled.bin
6	USRP B210 unit 2	SDR (OAI UE stack)	001010000000004	serial 3113EE5	pusch_dataset_b210_unit2_imsi004_51prb_100000labeled.bin

The IMSI is the class label, carried in every record. IMSI identifies the SIM, not the radio — the hardware IDs are what establish that each label corresponds to a distinct transmitter.

Radio configuration — identical for all six

gNB	OpenAirInterface, gnb.sa.band77.51prbs.x410.oaibox.conf
gNB host	NVIDIA DGX Spark (ARM64) + USRP X410
band	n77
carrier	51 PRB, 30 kHz SCS (numerology 1)
SSB	4008.00 MHz (ARFCN 667200)
PointA	4000.08 MHz (ARFCN 666672)
carrier centre	4009.26 MHz
TDD	DDSUU, 2.5 ms periodicity
CORESET0	10
min_rxtxtime	3
PUSCH allocation	fixed 51 PRB — the full carrier

Fixed by a gNB scheduler constant (FIXED_PUSCH_RB_SIZE = 51), so every record has rb_size == 51. Allocation width cannot leak class information.

At full carrier width rb_start is always 0 and only one UE can be granted uplink per slot, since rbStart + 51 > bwpSize rejects anything else. Every capture session therefore had exactly one device on the air — verified per session from the core's registration log.

Traffic: sustained bidirectional iperf3 (UDP, 5 Mbit/s each way) against the network's data endpoint, to drive real DTCH grants rather than signalling.

Record format

Binary, PUSC magic, format v5. 64-byte file header, then records with a 148-byte header plus payload. Key fields: rb_size (always 51), rb_start (read it — do not assume 0 in general), num_symbols, imsi (the label), timestamp_ns, iq_bytes, chest_bytes.

The gNB's own uplink channel estimate (chest) is stored alongside the IQ, same shape — enabling per-record link-quality metrics computed directly from the data, and equalisation experiments.

Reading the data

read_pusch_dataset.py is included — self-contained, numpy the only dependency, wire format defined inline. It doubles as the format specification.

python3 read_pusch_dataset.py scan pusch_dataset_b210_unit1_imsi001_51prb_100000labeled.bin
python3 read_pusch_dataset.py show pusch_dataset_rm520n_imsi007_51prb_100000labeled.bin 0

from read_pusch_dataset import iter_records
for meta, iq, chest in iter_records("pusch_dataset_pixel7_ue4_imsi010_51prb_100000labeled.bin"):
    # iq, chest: complex64, (num_symbols, 612)   612 = 51 PRB x 12 subcarriers
    label = meta["imsi"]

iter_records(..., with_payload=False) skips IQ decoding for metadata-only passes.

Per-device measurements

From the gNB's own MAC/PHY statistics during each session, collected before any container restart.

metric	RM520N A	RM520N B	Pixel UE4	Pixel UE5	B210 u1	B210 u2
UL SNR median	19.5 dB	19.5 dB	21.0 dB	20.0 dB	19.0 dB	20.0 dB
UL SNR range	15.5–24.5	16.5–22.5	1.5–24.0	14.0–35.5	16.5–23.5	8.0–25.0
UL BLER median	0.072	0.088	0.086	0.074	0.070	0.089
UL MCS table	1	1	1	1	0	0
UL rounds	109,751	105,484	121,578	107,626	171,227	122,600
UL errors	2	2	1	1	10	14
UL DTX	10	8	1	6	0	21
RSRP median	−71 dBm	−71 dBm	−72 dBm	−73 dBm	−80 dBm	−79 dBm
CCE fail (UL)	3	2	106	8	5	2
DRB TX/RX	1.00:1	1.00:1	1.00:1	1.00:1	1.00:1	0.99:1
UL failures	0	0	0	0	0	0

UL SNR medians span 19.0–21.0 dB across all six devices, and every DRB ratio is 1:1. This was achieved deliberately — see Transmit power calibration.

The occasional very low SNR minimum (Pixel UE4's 1.5 dB, B210 u2's 8.0 dB) is a single sample from a released RRC context printing final statistics, not a real channel event. Working ranges are ~6–13 dB wide.

Grant-size (num_symbols) distribution

device	13-sym (63,796 B)	3-sym (14,836 B)	12-sym (58,900 B)
RM520N-GL A	67,490 (67%)	29,278 (29%)	3,232
RM520N-GL B	66,106 (66%)	31,101 (31%)	2,793
Pixel 7 UE4	65,635 (66%)	32,232 (32%)	2,133
Pixel 7 UE5	72,415 (72%)	25,627 (26%)	1,958
B210 unit 1	61,377 (61%)	36,274 (36%)	2,349
B210 unit 2	77,472 (77%)	15,039 (15%)	7,489

Do not read this as a device signature. Three consecutive sessions of the same device (B210 unit 2, identical settings) produced 62% / 77% / 60% — a 17-point swing, wider than the 16-point spread across all six devices. Grant mix is session noise driven by traffic dynamics, not a stable per-device property. It is reported for transparency, not as a matched variable.

Transmit power calibration

The two B210s required explicit calibration. Uncalibrated (OAI's default --ue-txgain 0, i.e. full 89.75 dB) a B210 measured 41 dB UL SNR against ~20 dB for every commercial device — an offset a model separates on trivially. Worse, that reading reflected gNB receiver compression: the noise estimate was dominated by distortion products scaling with the signal, so the captured IQ contained the receiver's own non-linearity. A high SNR reading in this setup is not good news.

Final calibration, measured not assumed:

device	--ue-txgain	actual TX gain	resulting UL SNR median
B210 unit 1	26	63.75 dB	19.0 dB
B210 unit 2	29	60.75 dB	20.0 dB

Note --ue-txgain is attenuation from maximum in OAI (set_tx_gain(gain_range.stop() - tx_gain)), so higher means quieter.

The two units needed different settings because unit 2 measured ~3 dB louder at identical gain (different bench position and antenna coupling). Matching the setting and matching the received SNR were mutually exclusive; we matched received SNR, since that is what appears in IQ amplitude.

Consequence worth stating: the two B210s' power amplifiers therefore sit at slightly different backoff points (63.75 vs 60.75 dB), and PA non-linearity is part of what fingerprinting exploits. We judged a 3 dB gain-setting difference less harmful than a 3–4 dB received-SNR difference, which would be directly visible in IQ magnitude and perfectly correlated with class. A fully clean design would place both units at identical distance and identical gain — a bench-layout change, not a software one. This is a limitation, not a solved problem.

Device identity verification

IMSI comes from the SIM, so labelling by IMSI alone does not establish that two classes are two transmitters. Each pair was verified:

• RM520N-GL — AT+CGSN. IMEIs 868371056697259 / 868371056697580: distinct, sharing TAC 86837105 with near-sequential serials → same batch.
• Pixel 7 — IMEISV decoded by the core from NAS registration, needing no interaction with the handsets. 3536449372216689 / 3536449387387628: distinct, both TAC 35364493.
• USRP B210 — serial pinned in the launch command (--usrp-args serial=…) and confirmed in the softmodem log. 3113F40 / 3113EE5: distinct and adjacent → same batch. Both report the same USRP name (MyB210), so the name cannot distinguish them; only the serial can.

Two of three pairs are same-batch — the hardest case, since batch-mates have the closest component tolerances.

Confounds and limitations

Not fixable — inherent to the device classes

MCS table differs by class. Both B210s used MCS table 0; all four commercial devices used table 1. This follows from the OAI UE stack negotiating different capabilities than a commercial baseband, and it changes the modulation mapping — a genuine signal-level difference between the SDR and commercial classes. A model can separate SDR from commercial partly on this. The within-pair comparisons are unaffected.

Transmit chain implementation. The B210s run OAI's own PHY on general-purpose hardware — the same codebase as the gNB. If the SDR class proves easiest to identify, that is a plausible cause alongside RF hardware.

Measurement circumstances

RSRP spans 11 dB (−71 to −80 dBm), reflecting bench placement: modems and SDRs near the X410, phones further away. It did not propagate into UL SNR, which is matched. Per-record power normalisation is recommended preprocessing.

B210 PA backoff differs by 3 dB between units — see Transmit power calibration.

Pixel UE4 shows 106 UL CCE failures, an order of magnitude above the others. Its session was otherwise clean (1 UL error, DRB 1.00:1).

Session-to-session variance is substantial. Three identical B210 unit 2 sessions gave SNR medians of 18.0 / 20.0 / 15.0 dB. Single-session measurements here carry roughly ±2–3 dB of noise; treat the per-device table accordingly.

Excluded

Sessions discarded rather than published: uncalibrated B210 runs at 33–37 dB SNR; a session with two devices attached simultaneously (they contend for the whole carrier at 51 PRB); sessions where uplink was throttled by a host USB fault; and sessions where the traffic-carrying RNTI was never IMSI-labelled.

Intended use

Training and evaluating RF fingerprinting / device identification on real 5G NR uplink waveforms. The uplink-only, gNB-side design means the method transfers identically to SDR and commercial-baseband devices, requiring no instrumentation on the device.

Not for pairing, reciprocity, or key generation — uplink only, no matched downlink.

Suggested protocol: hold out by time (timestamp_ns) rather than sampling randomly, since adjacent records share channel state and random splits leak. Report within-pair accuracy separately from across-class accuracy — the pairs are the meaningful result. Apply per-record power normalisation.

Provenance

Captured 2026-08-05 on a self-hosted OpenAirInterface 5G SA network (DGX Spark + USRP X410) using the nr_pusch_capture plugin from the NVIDIA Sionna Research Kit. All six sessions used one gNB configuration and one gNB binary. Each session ran to 110,000 accepted slots and was trimmed to exactly 100,000 IMSI-labelled records, so classes are balanced. Every delivered file was verified to contain exactly one IMSI and rb_size == 51 for all 100,000 records.