ahancock516/oai-n77-pusch-20prb-fingerprint-dataset

5G NR PUSCH RF Fingerprinting — 6 Devices, 3 Same-Model Pairs (fixed 20 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 so that a model learns hardware differences rather than carrier or configuration differences.

Every record in this dataset is a fixed 20 PRB uplink allocation. That is the 20prb in the repository name, and it is a deliberate experimental variable, not just an implementation detail — see Fixed allocation width.

The point of this dataset is 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)

Distinguishing two units of the same model from the same batch is a much harder and more meaningful task than distinguishing a phone from an SDR. Every pair was verified to be physically distinct hardware using a hardware identifier, not assumed from a label — 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_100000labeled.bin
2	Quectel RM520N-GL unit B	embedded modem	001010000000006	IMEI 868371056697580	pusch_dataset_rm520n_imsi006_100000labeled.bin
3	Google Pixel 7 (UE4)	smartphone baseband	001010000000010	IMEISV 3536449372216689	pusch_dataset_pixel7_ue4_imsi010_100000labeled.bin
4	Google Pixel 7 (UE5)	smartphone baseband	001010000000008	IMEISV 3536449387387628	pusch_dataset_pixel7_ue5_imsi008_100000labeled.bin
5	USRP B210 unit 1	SDR (OAI UE stack)	001010000000001	serial 3113F40	pusch_dataset_b210_unit1_imsi001_100000labeled.bin
6	USRP B210 unit 2	SDR (OAI UE stack)	001010000000004	serial 3113EE5	pusch_dataset_b210_unit2_imsi004_100000labeled.bin

Three hardware classes — SDR, smartphone baseband, embedded modem — so the dataset supports both coarse (class) and fine (within-model) discrimination.

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

Radio configuration — identical for all six devices

gNB	OpenAirInterface, config gnb.sa.band77.51prbs.x410.oaibox.conf
gNB host	NVIDIA DGX Spark (ARM64) + USRP X410
band	n77
bandwidth	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 pattern	DDSUU, 2.5 ms periodicity (2 DL slots + 6 DL symbols, 2 UL slots + 4 UL symbols)
CORESET0	10
min_rxtxtime	3
PUSCH allocation	fixed 20 PRB for every record
PRACH DTX threshold	150
pMax / ssPBCH_BlockPower	26 / −20

PUSCH is pinned to exactly 20 PRB by a gNB scheduler patch (FIXED_PUSCH_RB_SIZE), so every record in the dataset has rb_size == 20. This removes allocation width as a confound entirely — a model cannot separate devices by how much bandwidth they were granted.

Fixed allocation width, and the 51 PRB companion

The fixed 20 PRB grant does two things at once, and it is worth separating them.

Within this dataset it is a control: allocation width cannot leak class information, because it is constant. Every device is fingerprinted from the same 20 PRB of occupied bandwidth (7.2 MHz at 30 kHz SCS), inside a 51 PRB carrier.

Across datasets it becomes an independent variable. The companion capture of the same six devices at fixed 51 PRB — the full carrier width — is now published:

ahancock516/oai-n77-pusch-51prb-fingerprint-dataset — same six physical devices, same band, same SSB/PointA placement, same TDD pattern, same gNB hardware. Only the uplink allocation width differs (612 subcarriers vs 240).

Because everything else is held constant, the pair isolates occupied-bandwidth effects on fingerprintability. Open questions that pairing makes answerable:

• Does more occupied bandwidth improve separability (more of each transmitter's frequency response is observed per record), or does it dilute the narrowband impairments that a 20 PRB view concentrates on?
• Do models trained at 20 PRB transfer to 51 PRB and vice versa, or is a fingerprint learned at one allocation width allocation-specific?
• Does the hardest comparison — the same-batch pairs — benefit more or less from the extra bandwidth than the easy across-class comparisons?

If you are building on this, treating allocation width as a studied axis rather than a fixed choice is likely more interesting than either dataset alone. Note that 51 PRB is also the practical ceiling for the B210 units over USB 3.0 (-E/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.

Traffic was driven to produce real DTCH uplink grants rather than registration-phase signalling. The four commercial devices ran sustained bidirectional iperf3 (UDP, 5 Mbit/s each way) against the network's data endpoint; the two B210s self-generated sufficient uplink traffic. See Confounds for the consequences.

Record format

Binary, PUSC magic, format v5. 64-byte file header, then fixed-size records with a 148-byte header followed by payload:

field	meaning
capture_idx	sequence number within the session
frame, slot	NR frame and slot the PUSCH occupied
rnti	gNB-assigned C-RNTI (changes across RRC re-establishment)
rb_size	always 20
rb_start	starting PRB — read this, do not assume 0
num_symbols	3, 12, or 13 (see the grant-size note below)
imsi	class label
timestamp_ns	wall-clock CLOCK_REALTIME
iq_bytes	raw uplink IQ
chest_bytes	the gNB's own uplink channel estimate, same shape as IQ

record_bytes is self-describing, so records can be walked without knowing num_symbols in advance.

Reading the data

read_pusch_dataset.py is included in this repository. It is self-contained — numpy is the only dependency, the wire format is defined inline, and there is nothing to clone or install. The file doubles as the complete format specification.

# summarise a file (streams; safe on multi-GB files)
python3 read_pusch_dataset.py scan pusch_dataset_b210_unit1_imsi001_100000labeled.bin

# dump one record's metadata plus IQ/chest shapes
python3 read_pusch_dataset.py show pusch_dataset_pixel7_ue4_imsi010_100000labeled.bin 0

from read_pusch_dataset import iter_records

for meta, iq, chest in iter_records("pusch_dataset_rm520n_imsi007_100000labeled.bin"):
    # iq, chest: complex64, shape (num_symbols, nb_re_per_sym)
    #            e.g. (3, 240) or (13, 240) -- 240 = 20 PRB x 12 subcarriers
    label = meta["imsi"]
    ...

iter_records(..., with_payload=False) skips IQ decoding when you only need metadata, which makes a full pass over a 1.8 GB file take seconds.

The channel estimate (chest) is included alongside the IQ. This makes it possible to compute per-record link-quality metrics directly from the data instead of trusting the aggregate values reported below, and to experiment with equalisation as a preprocessing step.

rb_start is not pinned. Single-UE sessions land at 0, but multi-UE sessions can place the block elsewhere in the carrier, so downstream code should read rb_start per record rather than assume a fixed absolute frequency bin.

Per-device measurements

Recorded from the gNB's own MAC/PHY statistics during each capture session, so that link conditions are visible rather than hidden. Read the Confounds section before training.

metric	RM520N A	RM520N B	Pixel UE4	Pixel UE5	B210 u1	B210 u2
UL SNR median	20.0 dB	20.0 dB	20.5 dB	20.0 dB	23.0 dB	24.0 dB
UL SNR range	17.0–23.5	17.5–26.5	15.5–25.0	16.5–22.5	19.5–27.5	21.0–29.0
UL BLER median	0.064	0.073	0.087	0.090	0.085	—
UL MCS table	1	1	1	1	0	0
UL rounds	118,630	104,229	157,293	113,049	111,631	128,158
UL errors	0	2	0	0	0	139
UL DTX	12	24	15	1	7	2,460
pucch0 DTX	8	13	4	2	10	6,306
RSRP median	−70 dBm	−75 dBm	−80 dBm	−82 dBm	−84 dBm	−86 dBm
RSRP range	−75..−68	−80..−73	−81..−78	−83..−78	−86..−81	−90..−83
DRB TX/RX ratio	1.05:1	2.46:1	1.01:1	1.00:1	2.92:1	0.54:1
CCE fail (UL)	3	3	10	6	3	—
UL failures	0	0	0	0	0	0

Grant-size (num_symbols) distribution

num_symbols determines record_bytes, so this is a direct readout of the grant sizes the scheduler assigned. It is the most important table here, because it is the one distributional difference that a model can exploit without examining the RF at all.

device	13-sym (25,108 B)	3-sym (5,908 B)	12-sym (23,188 B)
RM520N-GL A	65,237 (65%)	32,620 (33%)	2,143
RM520N-GL B	64,334 (64%)	33,563 (34%)	2,103
Pixel 7 UE4	63,064 (63%)	34,578 (35%)	2,358
Pixel 7 UE5	59,975 (60%)	37,589 (38%)	2,436
B210 unit 1	60,470 (60%)	37,610 (38%)	1,920
B210 unit 2	65,110 (65%)	32,312 (32%)	2,578

All six fall in a 60–65% band. This was deliberate: earlier capture attempts produced skews as extreme as 29% and 72% and were re-run rather than published, because a classifier can separate devices on symbol count alone. If you want it exactly balanced, stratify to the per-type minimum (59,975 of 13-sym and 32,312 of 3-sym → ~92,000 records per device).

Device identity verification

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

• Quectel RM520N-GL — AT+CGSN on each module. IMEIs 868371056697259 and 868371056697580: distinct, and sharing TAC 86837105 with near-sequential serials, i.e. same manufacturing batch.
• Google Pixel 7 — IMEISV decoded by the core from the NAS registration (Stored IMEISV in the NAS Context), requiring no interaction with the handsets. 3536449372216689 and 3536449387387628: distinct, both TAC 35364493. Independently corroborated by attaching both simultaneously and observing two entries in the AMF UE table.
• USRP B210 — USRP serial pinned in the launch command via --usrp-args serial=... and confirmed in the softmodem log. 3113F40 and 3113EE5: distinct and adjacent, i.e. same manufacturing batch. Both report the same USRP name (MyB210), so the name cannot distinguish them; only the serial can.

Two of the three pairs are same-batch, which is the hardest case: batch-mates have the closest component tolerances, so whatever separates them is closer to irreducible silicon variation than to model-level design differences.

Confounds and known limitations

Stated explicitly because a fingerprinting result is only meaningful if the model is learning hardware rather than measurement circumstance.

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 actual modulation mapping — so it is a genuine signal-level difference between the SDR class and the commercial classes, not merely a scheduling artifact. A model can separate SDR from commercial partly on this. The within-pair comparisons (B210 u1 vs u2, and each commercial pair) are unaffected.

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

Measurement circumstances

RSRP spans 12 dB across devices (−70 dBm to −86 dBm), reflecting physical placement: the modems and SDRs sit on the bench near the X410, the phones are further away. It did not propagate into UL SNR or grant mix, both of which are matched, but absolute received power is a per-device property here. Per-record power normalisation is recommended preprocessing and removes it.

UL SNR is matched to within 4 dB (20.0–24.0 dB medians) with heavily overlapping ranges. The B210s sit ~3–4 dB above the commercial cluster. This is the residue of an explicit calibration effort: both B210s' transmit gain was adjusted to bring their SNR into the commercial range. Uncalibrated, the B210 measured 41 dB — see below.

B210 unit 2 had an unstable link: 2,460 UL DTX and 6,306 pucch0 DTX against single/double digits for every other device, plus a 7 dB RSRP spread. Its records are valid captures, but its channel was more variable.

DRB TX/RX ratios vary by traffic pattern. The four commercial devices ran bidirectional iperf3 and sat at ~1:1. The B210s self-fed and landed off-balance in opposite directions (2.92:1 and 0.54:1). RM520N-GL unit B's 2.46:1 has a different cause: host-side USB packet loss upstream of the radio. Those packets never reached the air, so they cannot distort the transmissions that did occur — its grant-size distribution (64/34) is indistinguishable from unit A's (65/33).

Excluded / superseded

Earlier capture attempts were discarded rather than published:

• A B210 session at full transmit power measured 41 dB UL SNR with a 26 dB spread — this was the gNB receiver in compression, not a strong link. The noise estimate was dominated by distortion products scaling with the signal, so that IQ contained the receiver's own non-linearity. Discarded.
• Sessions with grant-size skews of 29% and 72% 13-symbol. Discarded.
• Sessions where uplink app traffic was throttled by a host USB fault. Discarded.

Consequently a high SNR reading in this setup is not good news — it can indicate front-end overload. Devices were deliberately calibrated down into a ~20 dB band.

Intended use

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

Not intended for pairing, reciprocity, or key-generation work: these are uplink captures only, with no matched downlink.

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

Provenance

Captured 2026-08-04 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. Per-device sessions were captured to 110,000 accepted slots and trimmed to exactly 100,000 IMSI-labelled records, so classes are balanced.

Label integrity: each session's records were verified to carry exactly one IMSI and rb_size == 20 for all 100,000 records before inclusion.