powder-lte-v1/docs/data_dictionary.md

# Data dictionary — requested fields vs. existing data

Stage 1 deliverable. For every field the proposal asks for, this table says whether it is
**available** in the existing raw archive (`data/raw/raw.7z`), **derivable** from it with the parsers in
`pipeline/parsers/`, or **needs new collection** on POWDER. "Where" names the raw file and the parser
output field. Provenance follows the project rule: `measured` = as captured, `derived` = computed
from measured data, `simulated` = not from real radios.

Status legend: **A** available (measured, populated) · **D** derivable (computed by a parser/summary) ·
**P** partially available (field exists but is unpopulated or unreliable — see note) · **N** needs new collection.

Verified against the captures on 2026-09-30 (tshark 4.6.4, srsRAN_4G fork commits `ed736f7` and `dd62a9b`).
Population statistics come from `inventory/catalog/catalog.json` (`extra.mr_nonzero_fraction`) and
`inventory/summaries/`.

## 1. Event data (Elasticsearch records)

| Requested field | Status | Where (raw → processed) | Notes |
|---|---|---|---|
| Event timestamp (UTC) | A | pcap frame time → `ts`, `ts_epoch_ns` in `*.s1ap_events`, `*.nas_events`, `*.rrc_events`; srsRAN log stamp → `*.log_events` | pcap and srsRAN log clocks are UTC and agree. Measurement-report and detector timestamps are on other, unaligned clocks (`ts_received_raw`, `ts_local_raw`, `clock` field). |
| Pseudonymous UE identifier | D | IMSI in Attach request / Identity response → `ue_id` (HMAC-SHA256, salt outside repo); M-TMSI → `tmsi_id` | Raw IMSI never written; validation checks for 15-digit leaks. RNTI (`rnti`) is kept as-is (temporary). |
| Cell identifier | A | SIB1 in UE/eNB MAC pcaps → `enb_id`, `cell_id_in_enb`, `cell_identity_hex`, `tac` (rrc_events); S1SetupRequest → `enb_id` (s1ap_events); PCI in measurement reports → `serving_pci`, `neighbors[].pci` | ECGI is not present in InitialUEMessage of these srsRAN captures (`s1ap.cell_ID` empty). |
| Procedure type | A | `procedure`, `message`, `pdu_type` (s1ap_events); `message` (nas_events, rrc_events) | All S1AP class-1/2 procedures named from TS 36.413 tables. |
| Outcome | A/D | `outcome` = request / indication / success / failure (s1ap, nas) | From PDU choice and NAS message type. |
| Failure cause | A | `cause_group`, `cause`, `cause_code` (S1AP Cause IE); `emm_cause`, `emm_cause_text`, `esm_cause` (NAS); `cause` in `log_events` procedure_failure records | Includes `unknown-targetID` handover-preparation failures (all two-ue and most single-ue runs). |
| Authentication events | A | NAS Authentication request/response/failure/reject, Security mode command/complete in s1ap_events and nas_events | Cellular (EPS-AKA) only. |
| Mobility events | A | HandoverRequired / HandoverPreparationFailure / HandoverCommand (s1ap_events); RRC MeasurementReport, RRCConnectionReconfiguration with `handover_command: true`, RRCConnectionReestablishmentRequest (rrc_events) | Benign runs contain completed S1 handovers between the two eNBs (HandoverRequired → HandoverCommand); fake runs contain HandoverPreparationFailure `unknown-targetID` towards the fake cell. The Polaris MME captures add X2 PathSwitchRequest. |
| Attach / service request / detach | A | NAS EMM messages (`nas_emm_message`), `Service reject` with `Implicitly detached` in fake runs | |
| Benign vs. malicious behaviour label (run level) | A | Directory name → `scenario_label` (benign / fake) on every record; `inventory/run_records/*.yaml` scenario.timeline | The activation instant of the fake eNB is not recorded. |
| Malicious-event label (frame level) | A (single-ue fake only) | `single-ue/fake/labels/labels.txt` → `frame_label` = `fbs-attributed` on `cn_real.s1ap_events` records | Frame ranges of the EPC S1AP capture attributed to the fake BS (handover attempts towards it). Not available for two-ue runs. |
| Detector alerts | A (separate) | `detector_outputs/**/*.csv` → `detector.*.ndjson`, `provenance: derived`, `record_type: detector_output`, `final_verdict_label` | Kept in their own index; never merged into the ground-truth timeline. Timestamps unaligned. |
| Application activity (web, file transfer, streaming) | N (LTE archive) / partial (Stage 3) | Stage 3 runs log ping and one HTTP GET per phase through the UE tunnel | No application logs in the LTE archive; MAC byte counts per second are the closest proxy (`*.mac_stats`, derived). |
| Node login / service-management events | N | — | Not collected. |
| Injected fault records (restart, config error, load, backhaul delay/loss) | N (LTE archive) / A (Stage 3) | Stage 3: `scenario.timeline` in every run record | LTE archive: only condition is FBS present/absent. |
| IMS events (registration, call setup/teardown) | N | — | Needs Kamailio IMS + Open5GS deployment. |
| RADIUS authentication / accounting | A (Stage 3) | FreeRADIUS 3.0.16 on the core node; `core/pcap/radius.pcap` → `*.radius_events.*` (Access-Request/Accept/Reject, Accounting Start/Interim/Stop with session id, octets, terminate cause, latency, pseudonymised user); `core/logs/freeradius.log` slice | Test clients (`radtest`, `radclient`) drive accept, wrong-password reject, unknown-user reject and accounting per round (`pipeline/collectors/radius_ctl.sh`). Separate from the 5G NAS authentication on N2. |

## 2. Network-performance data

| Requested field | Status | Where | Notes |
|---|---|---|---|
| Procedure counts per run | D | `inventory/summaries/procedure_counts.csv` | From s1ap_events per capture point. |
| Handover-preparation failure rate | D | `inventory/summaries/handover_outcomes.csv` (`prep_failure_rate`, causes) | HandoverPreparationFailure ÷ HandoverRequired. |
| Procedure durations (attach, service request) | D | `inventory/summaries/attach_service.csv` (`attach_ms_*`, `service_req_ms_*`) | Request/response correlated on ENB-UE-S1AP-ID within one capture. |
| Signal-quality change over time | D | measurement_reports + rrc MeasurementReport records (time series per UE) | See RF section for the underlying fields. |
| UL/DL goodput, RTT, jitter, packet loss | A (ZMQ runs) / N (LTE archive) | ZMQ runs: iperf3 JSON (`ran/app/iperf_*.json`), `ping -D` files, `curl -w` timings; POWDER core-sim: ping/HTTP only | Through a simulated radio link; absolute numbers describe the host, not a radio. LTE archive has none. |
| CQI | P (LTE) | `cqi` in measurement_reports | 0 in >99.9 % of LTE reports. Not exported by the ZMQ setup either (srsUE CSV has no CQI column); use DL SNR/MCS instead. |
| MCS (UL/DL) | A (ZMQ runs, simulated radio) | `ran.ran_metrics.ue_metrics` (`m_dl_mcs`, `m_ul_mcs`), gNB scheduler metrics in `ran.ran_metrics.gnb` | 1-second series per run from `data/raw/zmq_5g_sim`. LTE archive: always 0. |
| PRB / resource utilisation | A (ZMQ runs, simulated radio) | `ran.ran_metrics.gnb` scheduler cell metrics: `m_nof_prbs`, `m_pdsch_rbs_per_slot`, `m_pusch_rbs_per_slot`, `m_pdschs_per_slot`, `m_pucch_tot_rb_usage_avg` | ocudu `[METRICS]` lines. |
| HARQ / BLER | A (ZMQ runs, simulated radio) | `m_dl_bler`, `m_ul_bler` (UE CSV); `m_late_dl_harqs`, `m_late_ul_harqs`, `m_max_*_harq_delay` (gNB) | Vary under the srsUE AWGN / delay / RLF emulation scenarios. |
| DL SNR, RSRP, path loss, CFO, timing advance at the UE | A (ZMQ runs, simulated radio) | `m_dl_snr`, `m_rsrp`, `m_pl`, `m_cfo`, `m_ul_ta` in `ran.ran_metrics.ue_metrics` | Values reflect the ZMQ link and emulated channel, never air. |
| Throughput (MAC/RLC/user plane) | A/D (ZMQ runs) | iperf3 JSON in `ran/app/` (goodput, jitter, loss for UDP), `m_dl_brate`/`m_ul_brate` (UE), `m_total_dl_brate` (gNB), `gtpu_stats`, `mac_nr_stats` | iperf runs at a fixed target bitrate (10–20 Mbps) through the simulated link. |
| PUSCH/PUCCH SINR | A | `sinr_pusch_db`, `sinr_pucch_db` (measurement_reports) | Populated in 100 % of reports. |
| Power headroom, timing advance | A | `phr_db`, `ta_us` (measurement_reports) | Populated (>97 %). Timing advance is usually 0.6 µs (co-located SDRs). |
| Host metrics (CPU, memory, power) of the detector | A | `inference_performance.csv` → `detector.inference_performance.ndjson` | Only for the detector process, not for RAN/core nodes. |
| 5G performance KPIs | N | — | No 5G data at all in the archive. |
| PIM (passive intermodulation) | N | — | Not measurable from these data; requires calibrated RF test setup. Any modelled impairment must be labelled `simulated`. |

## 3. RF and spectral data

| Requested field | Status | Where | Notes |
|---|---|---|---|
| Serving-cell RSRP / RSRQ | A | `serving_rsrp_dbm`, `serving_rsrq_db` (measurement_reports, EPC side); `serving_rsrp_dbm`, `serving_rsrq_db` from RRC MeasurementReport (rrc_events, UE/eNB side, converted from TS 36.133 ranges) | Both sources present in every POWDER run that has UE attach. |
| Neighbour-cell RSRP / RSRQ / PCI | A | `neighbors[]` in measurement_reports and rrc MeasurementReport records | Neighbour is the second eNB (PCI 2 or 1). |
| SINR | A | `sinr_pusch_db`, `sinr_pucch_db` (uplink SINR measured at the eNB) | No downlink SINR at the UE. |
| Cell identifiers and timing | A | PCI, EARFCN, TAC, cell identity (SIB1), SFN/subframe (`sfn`, `subframe` in rrc_events), timing advance | |
| Broadcast parameters seen by the UE (barring, q-RxLevMin, bandwidth) | A | `cell_barred`, `q_rxlevmin` (SIB1), `dl_bandwidth_prb` (MIB) in rrc_events | Useful for FBS characterisation. |
| RSSI | P | `pusch_rssi_dbm`, `pucch_rssi_dbm` | Always 0. |
| Channel estimates, frequency offset, EVM | N | — | Not in any capture; needs PHY instrumentation (srsRAN PHY logs / OAI T-tracer). |
| GPS-tagged measurements for heatmaps | N (POWDER) / A (Polaris) | POWDER runs: no location data. Polaris Nemo `.nmf` → `gps` records (`lat`, `lon`, `altitude_m`) 1 Hz, with CELLMEAS RSRP/RSRQ in the same file | Nemo drive test (3 files, ~1 min each, Santa Clara CA, 14 Apr 2026). POWDER captures have no reconstructable locations → new collection with GPSD needed. |
| SDR IQ / power spectra, noise floor | N | — | No spectrum captures. |
| Antenna-port measurements (return loss, VSWR, cable loss, port power) | N | — | Requires RF instruments; not derivable from UE reports. |

## 4. 5G core data

| Requested field | Status | Where | Notes |
|---|---|---|---|
| S1AP (4G counterpart of NGAP) at the EPC | A | `cn_real/epc*.pcap`, `cn_fake/epc.pcap`, `enb_*/*_s1ap.pcap` → s1ap_events | Full S1AP with embedded NAS, one PDU per frame, no IP/SCTP layer (srsRAN raw writer). |
| Operator-scale S1AP (MME port mirror) | A (restricted) | `external/polaris/S1AP Data/*.pcap` (2 × ~0.3–0.7 GB, 31 Mar 2026) → same parser (handles Ethernet/VLAN/SCTP, bundled PDUs, Diameter/SGsAP alongside) | Third-party data with real subscriber identities; parse only with pseudonymisation on and keep raw restricted. |
| NAS at the UE | A | `ue*/ue*_nas.pcap` → nas_events | |
| RRC at UE and eNB | A | `*_mac.pcap` → rrc_events | |
| NGAP (gNB↔AMF), 5G NAS | A (Stage 3, simulated RAN) | `data/raw/powder_5g_sim/<run>/{core,ran}/pcap/n2_ngap.pcap` → `*.ngap_events.*` (message, procedure, outcome, cause, 5GMM/5GSM type and cause, PDU session id, pseudonymous SUPI) | Collected on POWDER with Open5GS 2.8.0 + UERANSIM 3.3.0 (`experiments/core-sim/`). Core signalling is real software; the radio side is simulated, so no RF fields. |
| SBI among AMF/SMF/NRF/AUSF/UDM/UDR (+SCP, PCF, BSF) | A (Stage 3) | `core/pcap/core_lo_cp.pcap` (loopback, plain HTTP/2) → `*.sbi_events.*` (service, resource, method, status, latency, consumer/producer, SCP leg) | Plain-text SBI in this deployment. Headers are only decodable when the HTTP/2 connection starts inside the capture, hence NFs are restarted at run start (recorded as ground truth) and `undecodable_headers` stubs mark any gap. |
| PFCP (SMF↔UPF) | A (Stage 3) | `core/pcap/core_lo_cp.pcap` → `*.pfcp_events.*` (message, cause, SEID, F-TEID, UE IP, latency) | |
| GTP-U user plane (N3) | A/D (Stage 3) | `{core,ran}/pcap/n3_gtpu.pcap` → `*.gtpu_stats.*` per-second bytes/packets per TEID and direction (derived) | Raw user-plane pcap kept separate from control plane. |
| NF application logs | A (Stage 3) | `core/logs/<nf>.log` slices → `*.open5gs_log.<nf>` (component, level, event class, pseudonymised SUPI) | Local-time stamps converted with the recorded node offset. |
| Injected faults with ground truth | A (Stage 3) | `run_record.yaml` `scenario.timeline` (`fault_start`/`fault_end`, phase tags) from `pipeline/collectors/fault_inject.sh` | NF restart, NF stop for N s, netem delay/loss on N2/N3. |
| Application workload at the UE | A (Stage 3, minimal) | `ran/app/ping_*.txt`, `ran/app/http.txt` (ping -D, curl timing through `uesimtun0`) | Ping/HTTP only; goodput tests (iperf) not yet run. |

## Cross-cutting fields required by the data rules

| Field | Status | Where |
|---|---|---|
| Provenance label | D | `provenance` on every processed record and summary row; run records list outputs with labels |
| Run identifier | D | `run_id` (`pfl-<variant>-<scenario>-exp-NN`, `detector_outputs-server-data-exp-01`, `polaris-*`) on every record; `inventory/run_records/<run_id>.yaml` |
| Software versions | A | srsRAN commit + branch from log start lines → run record `software_versions`; configs themselves are **not** in the archive |
| Capture points | A | directory names → `capture_point`; per-file sha256 in run record |
| Clock-sync information | P | pcap/log clocks consistent (UTC); MR/detector clocks unverified → `clock_sync` note in run record and `notes` in `inventory/catalog/runs.csv` |
| Train/test split by whole runs | A | `powder_fbs_lte/single-ue/train_test_split.json` → `split` column in `runs.csv` and run records (single-ue only; two-ue has no split file) |
| Dropped packets/samples | N | not recorded at capture time; `validation/report.md` compares parsed vs. raw counts instead |