ADAM-V2X paper evaluation dataset
Release preparation date: 2026-08-24
This dataset contains the Raspberry Pi multi-run measurements used to reproduce
the paper's QoS, detection, CPU-time, and separately profiled memory results.
Each reported configuration uses 20 independent trials.
Archives
adam-v2x-paper-data-v1-qos-mode0.tar.gz: continuous-uniform workloads
(Mode 0), including the clean baseline, Native, CoDel, and ONNX profiles.adam-v2x-paper-data-v1-qos-mode1.tar.gz: single-pulse workloads
(Mode 1), including Native at 10% and ONNX at all evaluated rates.adam-v2x-paper-data-v1-qos-mode2.tar.gz: periodic-pulsing workloads
(Mode 2), including Native at 10%, ONNX at all evaluated rates, and the
FSM-only and Static-100% ablations at 0.1% and 10%.adam-v2x-paper-data-v1-memory.tar.gz: logging-disabled peak-RSS records
fromoutputs/pure_memory_runs/.
The three QoS archives contain only run_01 through run_20. Local smoke-test
runs such as run_991 through run_996 are excluded.
Workload-rate semantics
The requested attack rate is the malicious fraction over the complete session.
Temporal modes concentrate that global packet budget into their active windows:
- Mode 0: continuous uniform injection.
- Mode 1: one pulse occupying 20% of the session.
- Mode 2: periodic windows occupying 50% of the session.
CSV formats
QoS files begin with a metadata line followed by:
packet_id,is_malware,was_dropped,latency_ns
The peak_rss_kb value embedded in a packet-logging QoS file is not the RAM
measurement reported by the paper because CSV buffering affects that process.
Paper RAM values must be derived from the logging-disabled memory archive. Each
memory file contains one line:
# METADATA: peak_rss_kb=<value>
Mode 1 and Mode 2 Native RAM values were not measured or reported; those cells
are intentionally absent from the memory dataset.
Derived timeline statistics
timeline_representative_statistics.csv is a diagnostic table derived from
the representative (median-selected) Mode-1 and Mode-2 runs. Its
region_packet_samples column is the number of received packet records inside
the named timeline region; it is not an attack rate, malicious-packet count, or
inspection-sampling rate. global_attack_rate_pct gives the requested malicious
fraction over the complete session, and region_definition identifies the
packet-ID window used for each statistic.
scheduled_attack_rate_within_region_pct records the corresponding scheduled
intensity inside that region (50% for the Mode-1 pulse and 20% for the Mode-2
attack windows when the session-global rate is 10%). This is workload injection,
not the admission filter's inspection-sampling rate.
timeline_manuscript_summary.csv is the unambiguous four-row subset quoted by
the manuscript: Mode 1 and Mode 2, Native and ONNX, raw receiver latency inside
the attack region. The manuscript uses only mean_ms and p99_ms from these
four rows. Clean-region and rolling-window rows in the diagnostic table are
retained for audit and figure interpretation, not as additional paper results.
Completeness notes
All expected scenario/trial files are present: 240 Mode-0 QoS files, 120 Mode-1
QoS files, 200 Mode-2 QoS files, and 520 memory records.
Seven QoS files contain fewer than 1,000,000 received packet rows and are
retained as measured rather than silently repaired or removed:
- Mode 0, run 01, CoDel at 1.0%: 999,999 rows (1 packet not received).
- Mode 1, run 02, Native at 10.0%: 999,892 rows (108 packets not received).
- Mode 1, run 03, Native at 10.0%: 999,882 rows (118 packets not received).
- Mode 1, run 04, Native at 10.0%: 999,398 rows (602 packets not received).
- Mode 1, run 05, Native at 10.0%: 999,340 rows (660 packets not received).
- Mode 1, run 06, Native at 10.0%: 999,529 rows (471 packets not received).
- Mode 1, run 08, Native at 10.0%: 999,438 rows (562 packets not received).
The archive checksums are recorded in SHA256SUMS and the per-file inventory is
recorded in FILE_INVENTORY.tsv.