powder-lte-v1/docs/README.md

Stage 1 — Inventory of existing datasets

What is in the raw archive (raw.7z), how it is organised, what is wrong with it, and how it was catalogued. The raw archive itself is not distributed; only the metadata and processed records are.

Contents of the archive

Dataset	Runs	Files	Size	What it is
powder_fbs_lte/single-ue	20 benign + 17 fake	438	4.2 GB	srsRAN_4G LTE on POWDER: one srsUE, one real eNB+EPC, a second eNB that is a fake base station in fake runs
powder_fbs_lte/two-ue	15 benign + 11 fake	379	3.4 GB	same with two srsUEs
detector_outputs	17 (+1 standalone server-data)	55	1.9 MB	CSV outputs of the FBS detector for a subset of two-ue runs
external/polaris/S1AP Data	2	1 547	3.1 GB	third-party operator MME S1AP port-mirror captures ("VDF-LTE DC1/DC2", 31 Mar 2026), full pcapng + 3 000-packet splits
external/polaris/RE_ …	3	6	0.7 MB	Keysight Nemo Outdoor drive-test logs (.nmf) with GPS, plus RRC/NAS pcaps converted by the included nmf_to_pcap.py

Totals: 2 427 files, 10.7 GB unpacked, 69 catalogued runs. Exact per-file and per-run metadata: catalog/files.csv, catalog/runs.csv, catalog/duplicates.csv, catalog/catalog.json.

Layout of one POWDER run

powder_fbs_lte/<variant>/<scenario>/exp-N/
  cn_real/   epc.pcap | epc_capture.pcap    S1AP at the real EPC (srsEPC raw S1AP writer, DLT USER 3)
             epc_measurement_reports.json   NDJSON: UE measurement reports relayed to the EPC (RSRP/RSRQ/SINR/PHR/TA…)
             epc.xml                        (exp-1 only) tshark PDML export of epc.pcap
  cn_fake/   epc.pcap                       S1AP at the fake EPC (fake runs; a few benign runs carry a stale copy)
  enb_real/  enb[1]_s1ap.pcap, enb[1]_mac.pcap, enb[1].log     real eNB: S1AP, MAC-LTE (RLC/PDCP/RRC/NAS inside), srsENB log
  enb_fake/  enb2_s1ap.pcap, enb2_mac.pcap, enb2.log           second eNB (fake BS in fake runs; attached to the real EPC in benign runs)
  ue|ue1|ue2/ ue*_mac.pcap, ue*_nas.pcap, ue*_mac_nr.pcap, ue.log   srsUE: MAC-LTE, NAS, (empty NR MAC), UE log (single-ue exp-1..7 only)

<variant> is single-ue or two-ue; <scenario> is benign or fake and is the run-level ground-truth label. single-ue/fake/labels/labels.txt adds frame-level ground truth for the single-ue fake runs: per experiment, 1-based frame-number ranges of the cn_real S1AP capture whose messages are attributed to the fake base station (in practice the HandoverRequired → HandoverPreparationFailure pairs towards the fake cell). parse_run copies these onto the cn_real.s1ap_events records as frame_label (fbs-attributed / unlabelled). No such file exists for two-ue runs. Note that exp1's ranges refer to the original single-ue capture, not the contaminated copy now in exp-1/cn_real (see problem 6). run_id = pfl-<variant>-<scenario>-exp-NN (zero-padded). single-ue/train_test_split.json defines a run-level train/test split for single-ue runs (14+10 train, 6+7 test); there is none for two-ue.

Software: srsRAN_4G fork with network-side FBS-detection patches. January 2026 runs use commit ed736f7 (branch feature/s1-handover-support), March/April 2026 runs use dd62a9b (fbs-location-simplified). Configuration files (enb.conf, enb2.conf, ue.conf) are referenced in the logs but are not in the archive.

Typical run: S1 setup, UE attach(es) with EPS-AKA, periodic RRC MeasurementReports, repeated HandoverRequired → HandoverPreparationFailure (unknown-targetID, because the target eNB is not known to the MME), UE context release on inactivity. In fake runs the UE additionally camps on the fake cell (eNB id 6592, PCI 2) and gets Service reject (Implicitly detached) from the fake EPC.

Known problems (all handled by the parsers, all flagged in catalog/runs.csv)

1. Clocks. pcap and srsRAN log timestamps are UTC and agree. received_at in the measurement-report files and the detector CSV timestamps come from other hosts, carry no timezone, and are offset from the pcaps by hours in several runs. Parsed records keep them raw (ts_received_raw, ts_local_raw) with a clock tag; do not join on them until the offset is established per run.
2. Appended measurement-report files. Some epc_measurement_reports.json span >24 h and contain reports from earlier runs (e.g. two-ue/benign/exp-5: 2 493 lines over 18 h). Filter by S1AP-id/RNTI or by time window after fixing the clock offset.
3. Invalid JSON. January runs have "mcc":,"mnc": (empty values). The parser repairs these lines (→ null) and sets parse_repaired; two runs also contain truly unparseable lines (mr_unparseable).
4. Unpopulated fields. In the measurement reports cqi, ri, pmi, dl_mcs, ul_mcs, dl_bler, ul_bler, dl_bitrate, ul_bitrate, dl_pkts, ul_pkts, pusch_rssi_dbm, pucch_rssi_dbm, turbo_iters are 0 in

   99.9 % of records. Only RSRP/RSRQ (serving + neighbour), PUSCH/PUCCH SINR, PHR and TA are real.

5. Empty captures. Every *_mac_nr.pcap is a 24-byte header with no packets.
6. Byte-identical files across runs (catalog/duplicates.csv): two-ue/benign/exp-14|15|16/ue1/* are the same files; single-ue/fake/exp-1/cn_real/* equals two-ue/fake/exp-1/cn_real/* (single-ue/fake/exp-1 cn_real is therefore contaminated — its eNB logs are from January but the EPC files are from March); cn_fake/epc.pcap in benign two-ue runs 11–16 are stale copies of fake-run files. Treat those files as unreliable and never count the duplicates twice in train/test.
7. Missing capture points. catalog/runs.csv column missing_expected lists absent files per run (e.g. two-ue/benign/exp-6 has no cn_real and no enb_real S1AP pcap).
8. Secrets in logs. srsENB warning lines dump PDCP integrity keys as hex. The log parser drops every continuation line; raw logs are never copied into the processed records.
9. Third-party data. The Polaris MME captures contain real IMSIs/TMSIs of an operator network. Only parse with the pseudonymisation salt in place (generated on first run, never distributed) and keep the raw files out of any shared location. The Nemo RRC→pcap conversion decodes PCCH paging correctly but DCCH messages are partly misdecoded, and the converter writes the handset's local time (UTC-7) as if it were UTC, so those pcaps are 7 h early; nmf_gps applies the #UT offset from the NMF header and is correct. Treat the converted RRC records as indicative only. The two MME captures cover only ~2 minutes each (31 Mar 2026 08:51–08:54 UTC).

How the inventory was produced

The raw archive was extracted and catalogued (sha256, capinfos, log and measurement-report summaries), each POWDER run got a retro-fitted run record, the captures were parsed into NDJSON event records, derived summaries were computed, and parsed records were checked against raw packet counts.

Outputs:

• data_dictionary.md — every requested field marked available / derivable / needs new collection.
• catalog/ — file and run catalog.
• run_records/ — one retro-fitted run record per POWDER run.
• summaries/ — handover outcomes, attach/service-request durations, radio-quality statistics, cells seen (derived).
• processed/ (in the tarball) — structured LTE event records (NDJSON).