Lystea/CICIOT2023-PARQUET

CICIoT2023 — ipfixprobe flow records (Parquet)

1,479,074,715 bidirectional network flows re-exported from the raw PCAPs of CICIoT2023 (Canadian Institute for Cybersecurity, University of New Brunswick) with ipfixprobe 5.7.0, stored as 308 Parquet files (~19.4 GB) covering 33 attack classes + benign traffic from the 105-device IoT testbed.

The original dataset ships ~587 GB of PCAPs and CSV features computed with a closed pipeline. This conversion provides an alternative featureization from an open, reproducible flow exporter: every record includes per-packet size/timing sequences for the first 30 packets (suitable for sequence models such as the CESNET "30 packets" family), size/inter-packet-time histograms, and TLS/QUIC handshake metadata — the same features a production ipfixprobe sensor emits, so models trained here transfer to live capture without a feature-skew confound.

Files

One Parquet file per source PCAP, named {Class}__{pcap-stem}.parquet (e.g. DDoS-TCP_Flood__DDoS-TCP_Flood11.parquet ← DDoS-TCP_Flood/DDoS-TCP_Flood11.pcap). Labels are per source capture, identical to the original dataset's ground truth, and repeated in the label_class column. manifest.json / hf-manifest.json record per-file conversion provenance (source bytes, outcome, timing).

Schema (35 columns)

Group	Columns	Notes
Flow key	src_mac, dst_mac, src_ip, src_port, dst_ip, dst_port, protocol	bidirectional (biflow), forward = initiator
Counters	packets, packets_rev, bytes, bytes_rev, tcp_flags, tcp_flags_rev	per direction
Timing	ts_start, ts_end, duration	epoch seconds (float)
Per-packet info (pstats)	ppisizes, ppitimes, ppiflags, ppidirs	first 30 packets; sizes are transport payload bytes (not frame size); times are inter-packet gaps in ms (first = 0); directions encoded +1/-1; zero-payload packets (bare ACKs) excluded from PPI but still counted in packets
Histograms (phists)	sphistsize, sphistipt, dphistsize, dphistipt	8 log-scale bins per direction; sizes in bytes (0–15 … >1024), IPT in ms
TLS / QUIC	tlssni, tlsalpn, tlsversion, tlsja3, tlsexttype, tlsextlen, quicsni	empty for non-TLS/QUIC flows
Provenance / label	source_dataset, source_file, label_class, end_reason	end_reason is unpopulated (null)

Flow definition: ipfixprobe cache with CESNET's documented timeouts — active 300 s, inactive 65 s. Process plugins: pstats, phists, tls, quic.

Class distribution

Class	Flows		Class	Flows
DDoS-TCP_Flood	371,310,722		Recon-PortScan	193,057
DDoS-RSTFINFlood	332,327,358		DDoS-ICMP_Flood †	183,642
DDoS-PSHACK_Flood	327,190,900		Mirai-greeth_flood	174,812
DDoS-SYN_Flood	326,685,573		Recon-OSScan	172,051
Mirai-greip_flood	72,345,759		DDoS-UDP_Fragmentation	146,859
DDoS-ACK_Fragmentation	12,291,718		Mirai-udpplain	140,159
DoS-SYN_Flood	12,246,404		DDoS-SlowLoris	122,119
DoS-TCP_Flood	8,652,190		DNS_Spoofing	81,951
DDoS-SynonymousIP_Flood	7,903,290		MITM-ArpSpoofing	73,066
DoS-UDP_Flood	3,842,424		DictionaryBruteForce	7,877
DDoS-UDP_Flood	791,303		SqlInjection	5,389
Recon-HostDiscovery	426,062		CommandInjection	4,010
VulnerabilityScan	403,524		XSS	3,479
DDoS-HTTP_Flood	400,951		BrowserHijacking	2,807
Benign_Final (benign)	358,565		Backdoor_Malware	2,455
DDoS-ICMP_Fragmentation	298,780		Recon-PingSweep	1,706
DoS-HTTP_Flood	282,398		Uploading_Attack	1,355

Class imbalance is extreme (floods dominate ~98% of flows) — this reflects the source captures, not a sampling choice. Weight or subsample accordingly.

Known gaps and caveats

• 308 of 309 source PCAPs are included. DDoS-ICMP_Flood/DDoS-ICMP_Flood.pcap deterministically crashes ipfixprobe 5.7.0 and is absent († the DDoS-ICMP_Flood class retains the flows from its other 26 capture parts).
• Labels are per capture file. As in the original dataset, any incidental background traffic inside an attack capture inherits the attack label.
• The feature set differs from the official CICIoT2023 CSVs — results are not directly comparable with papers using those features.
• Source PCAPs were fetched from the public Hugging Face mirror bencorn/CIC-IoT-2023 and verified by exact size (HF tree API) and PCAP magic bytes; CIC publishes no official checksums.

Reproducibility

Exporter: ipfixprobe 5.7.0 (Alpine/musl, aarch64 build), plugins and parameters as above (pstats: max 30 packets, min length 1, skip duplicates, exclude zero-payload; phists: 8 log bins). Re-exporting with a different exporter build or different cache timeouts changes what counts as a flow and will shift every downstream feature — pin accordingly.

License and citation

The underlying traffic captures belong to the Canadian Institute for Cybersecurity and are redistributed here in derived form for research use under the CIC dataset terms. If you use this data, cite the original paper:

@article{neto2023ciciot,
  title   = {{CICIoT2023}: A Real-Time Dataset and Benchmark for Large-Scale Attacks in {IoT} Environment},
  author  = {Neto, Euclides Carlos Pinto and Dadkhah, Sajjad and Ferreira, Raphael and Zohourian, Alireza and Lu, Rongxing and Ghorbani, Ali A.},
  journal = {Sensors},
  volume  = {23},
  number  = {13},
  pages   = {5941},
  year    = {2023},
  doi     = {10.3390/s23135941}
}

Please also credit CESNET/ipfixprobe for the flow exporter.