Datasets:

DeepTempo
/

cic-ids-2017-flowprep

license: other
license_name: cic-ids-2017-research-use
license_link: https://www.unb.ca/cic/datasets/ids-2017.html
pretty_name: CIC-IDS-2017 — Canonical NetFlow (flowprep)
language:
  - en
task_categories:
  - tabular-classification
tags:
  - network-security
  - cybersecurity
  - netflow
  - flow
  - intrusion-detection
  - nids
  - canonical-schema
  - flowprep
  - deeptempo
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/cic-ids-2017-canonical.parquet
dataset_info:
  config_name: default
  features:
    - name: timestamp
      dtype: int64
    - name: src_ip
      dtype: string
    - name: dest_ip
      dtype: string
    - name: src_port
      dtype: int32
    - name: dest_port
      dtype: int32
    - name: fwd_bytes
      dtype: int64
    - name: bwd_bytes
      dtype: int64
    - name: fwd_pkts
      dtype: int64
    - name: bwd_pkts
      dtype: int64
    - name: flow_dur
      dtype: float64
    - name: protocol
      dtype: int32
    - name: attack
      dtype: string
    - name: label
      dtype: string
  splits:
    - name: train
      num_examples: 101094

CIC-IDS-2017 — Canonical NetFlow (flowprep)

A ~101k-row slice of the CIC-IDS-2017 intrusion-detection dataset, normalized into a single clean, typed, unit-normalized canonical NetFlow parquet table by flowprep — the open-source flow canonicalization tool DeepTempo runs in production ahead of inference.

This is a conversion proof-point: it shows what a real, messy research dataset looks like after flowprep canonicalize resolves its vendor column names, infers its units, and types its columns. The output is ordinary parquet — pandas/polars/DuckDB/Spark read it natively, the file format is the API.

⚠️ This is a small demonstration slice, not the full CIC-IDS-2017 dataset. For research, obtain the complete dataset from UNB CIC and cite accordingly (see Citation).

Why this exists

Every network-ML project and SOC pipeline starts by solving the same unglamorous problem: flow data arrives with different column names, duration units, and timestamp encodings. The same field is src_ip, Source IP, srcaddr, or ipv4_src_addr; a duration might be seconds, ms, µs, or ns with nothing telling you which. flowprep maps 100+ column spellings onto one canonical schema and normalizes units, so labeled research data stays labeled and is ready to train on immediately. This dataset is that step, done once, published.

How it was produced

Fully reproducible from the sample bundled in the flowprep repo (no network access required):

git clone https://github.com/DeepTempo/flowprep
cd flowprep
cargo build --release

# aliased CIC columns (total_fwd_pkts, datetime timestamps) -> canonical schema
./target/release/flowprep canonicalize \
    examples/cic2017_sample.parquet \
    cic-ids-2017-canonical.parquet
# -> Wrote 101094 flows

flowprep resolved the source's CIC-style aliases (e.g. total_fwd_pkts → fwd_pkts, total_bwd_pkts → bwd_pkts), converted the typed-datetime timestamp to int64 epoch microseconds, kept flow_dur in float64 seconds, and passed the ground-truth label columns through unchanged. The alias map is not hard-coded — it is loaded at compile time from flowprep's declarative schemas/netflow/v1/schema.json, the same artifact DeepTempo's production ingestion uses. Output is single-row-group ZSTD-compressed parquet.

Schema

field	type	unit / encoding	description
`timestamp`	int64	epoch microseconds	flow start time
`src_ip`	string		source IP
`dest_ip`	string		destination IP
`src_port`	int32		source L4 port
`dest_port`	int32		destination L4 port
`fwd_bytes`	int64	bytes	source→destination bytes
`bwd_bytes`	int64	bytes	destination→source bytes
`fwd_pkts`	int64	packets	source→destination packets
`bwd_pkts`	int64	packets	destination→source packets
`flow_dur`	float64	seconds	flow duration
`protocol`	int32	IANA number	null in this slice — the CIC sample carries no protocol field; kept for canonical-schema fidelity
`attack`	string	`benign` / `attack`	ground-truth label (string)
`label`	string	`0` / `1`	ground-truth label (binary, string-encoded)

attack and label are two encodings of the same binary ground truth and are preserved exactly as flowprep emitted them (label passthrough). This slice does not carry per-attack subtype labels (DoS, PortScan, …).

Statistics

Rows: 101,094 flows
Class balance: 81,171 benign (80.29%) / 19,923 attack (19.71%)
Unique source IPs: 4,656 · unique destination IPs: 7,254
Timestamp range (as stored, passed through from source): 2017-03-07 → 2017-07-07 UTC. CIC-IDS-2017 was captured over a work week in July 2017; the exact values here are the source sample's timestamps, unchanged by flowprep.
flow_dur: 0 – 120.0 s (mean 14.84 s)
fwd_bytes: 0 – 1,034,689 (mean 527) · bwd_bytes: 0 – 538,000,000 (mean 16,097)
fwd_pkts: 1 – 180,725 (mean 9.4) · bwd_pkts: 0 – 242,974 (mean 10.5)
File: data/cic-ids-2017-canonical.parquet (~1.7 MB, ZSTD, single row group)

Usage

import pandas as pd
df = pd.read_parquet("hf://datasets/DeepTempo/cic-ids-2017-flowprep/data/cic-ids-2017-canonical.parquet")
# that's the whole integration — no bindings, no client library

# binary label for training/eval
y = (df["label"] == "1").astype(int)

# or with the datasets library
from datasets import load_dataset
ds = load_dataset("DeepTempo/cic-ids-2017-flowprep", split="train")

import duckdb
duckdb.sql("SELECT attack, count(*) FROM 'data/cic-ids-2017-canonical.parquet' GROUP BY attack")

Intended use

A small, clean, labeled flow table for prototyping NIDS / flow-classification models and for reproducible evaluation (the canonical schema removes preprocessing variance — a documented source of up to ~40% metric swing across NIDS papers).
A worked example of canonical NetFlow for the new generation of network foundation models, which expect one stable, unit-normalized schema.
A reference for what flowprep output looks like before you point it at your own pcap / nfdump / OCSF / CSV sources.

Limitations & honest notes

Not the full dataset. A ~101k-flow demonstration slice; not class-complete and not a substitute for the official CIC-IDS-2017 release.
Binary labels only. attack/label are benign-vs-attack; no per-attack subtype is present in this slice.
protocol is null (the source sample has no protocol field); the column is retained only for canonical-schema fidelity.
Timestamps are passed through from the source unchanged; treat them as relative ordering, not authoritative wall-clock capture times.
Known CIC-IDS-2017 caveats apply. The original dataset has documented labeling and generation issues discussed in the literature; do not treat results on this slice as production detection performance.

License & attribution

The canonical parquet artifact and this dataset card are released by DeepTempo; the flowprep tool that produced it is Apache-2.0.

The underlying flow data derives from CIC-IDS-2017, © Canadian Institute for Cybersecurity (CIC), University of New Brunswick. It is redistributed here as a small conversion example, with attribution and citation as required by CIC. For research use, obtain the full dataset from UNB CIC and cite the paper below. If you are a rights holder and have concerns about this slice, please open an issue on the flowprep repo.

Citation

If you use this data, cite the original CIC-IDS-2017 paper:

@inproceedings{sharafaldin2018toward,
  title     = {Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Characterization},
  author    = {Sharafaldin, Iman and Lashkari, Arash Habibi and Ghorbani, Ali A.},
  booktitle = {Proceedings of the 4th International Conference on Information Systems Security and Privacy (ICISSP)},
  year      = {2018}
}

And, optionally, the tool used to canonicalize it:

@software{flowprep,
  title  = {flowprep: network telemetry to ML-ready canonical NetFlow parquet},
  author = {DeepTempo},
  url    = {https://github.com/DeepTempo/flowprep},
  year   = {2026}
}

Acknowledgments

Canonical Institute for Cybersecurity (UNB) for CIC-IDS-2017.
Built with flowprep by DeepTempo.