AINovice2005/cicflow-ids-multiclass

CICFlow Multiclass Intrusion Detection Dataset

Overview

This dataset provides a multiclass network intrusion detection (IDS) benchmark derived from CICFlowMeter flow-level features. It is designed for attack-type classification, robust IDS research, and interpretable security modeling.

Each network flow is labeled as either benign or one of nine attack categories, following a standard IDS taxonomy.

The dataset is published in Hugging Face datasets format with explicit schemas, clean preprocessing, and reproducible splits.

---

Task

Multiclass classification

Given a network flow represented by CICFlowMeter features, predict the attack category.

---

Label Taxonomy

The dataset uses the following 10-class label space:

Label ID	Name	Description
0	benign	Normal network traffic
1	analysis	Port scans, probing, and analysis activity
2	backdoor	Backdoor and remote access behavior
3	dos	Denial-of-Service attacks
4	exploits	Exploitation of vulnerabilities
5	fuzzers	Fuzzing and malformed input attacks
6	generic	Generic attack traffic
7	reconnaissance	Reconnaissance and information gathering
8	shellcode	Shellcode execution attempts
9	worms	Worm propagation traffic

Additional label fields

• label → multiclass label (0–9)
• is_attack → binary indicator (1 if label ≠ 0, else 0)

This allows both multiclass IDS and binary detection experiments without reprocessing.

---

Dataset Structure

DatasetDict({
  train,
  validation,
  test
})

Each split contains records with the following schema:

flow_id: string
features: dict[str, float]
semantic_flags: dict[str, int]
label: ClassLabel (0–9)
is_attack: int (0/1)

---

Feature Description

1. Raw Numeric Features (features)

The features field contains CICFlowMeter-derived flow statistics, including:

Traffic volume & direction

• total_fwd_packets, total_bwd_packets
• total_length_of_fwd_packets
• total_length_of_bwd_packets
• down_up_ratio

Packet size statistics

• packet_length_min, packet_length_max
• packet_length_mean, packet_length_std, packet_length_variance
• Forward and backward packet length statistics

Timing & inter-arrival times

• flow_duration
• flow_iat_mean, flow_iat_std, flow_iat_max, flow_iat_min
• Forward and backward IAT statistics
• active_*, idle_* metrics

Rate-based features

• flow_bytes_per_s
• flow_packets_per_s
• fwd_packets_per_s, bwd_packets_per_s

TCP flag counters

• syn_flag_count, ack_flag_count, rst_flag_count
• fin_flag_count, psh_flag_count, urg_flag_count
• cwr_flag_count, ece_flag_count

Bulk and subflow statistics

• *_bulk_* features (conditionally emitted by CICFlowMeter)
• subflow_fwd_*, subflow_bwd_*

Note: Some CICFlowMeter features are conditionally emitted. Missing or undefined numeric values were filled with 0.0, which semantically indicates absence of that behavior.

---

2. Semantic Flags (semantic_flags)

To support interpretable and rule-based IDS, each flow includes deterministic semantic indicators:

Flag	Meaning
high_packet_rate	Very high packets per second
one_way_traffic	Forward-only traffic (no response)
syn_ack_imbalance	Large SYN/ACK imbalance
uniform_packet_size	Low packet size variance
long_idle_c2	Long idle periods (possible C2 behavior)

These flags are:

• Derived deterministically from numeric features
• Stable across dataset versions
• Suitable for rule-based or hybrid ML systems

---

Preprocessing Summary

The dataset was produced using a fully deterministic pipeline:

• Column name normalization
• Conversion of NaN / ±Infinity → 0.0
• Explicit label normalization (string → numeric)
• No row drops based on missing numeric values
• Stratified train/validation/test splits
• Explicit Hugging Face feature schemas

No synthetic balancing, oversampling, or augmentation was applied.

---

Class Distribution & Imbalance

The dataset is highly imbalanced, reflecting real-world network traffic:

• Benign traffic dominates
• Some attack classes are rare

This is intentional and realistic.

Users are strongly encouraged to:

• Use macro / weighted F1
• Inspect per-class recall
• Apply class weighting where appropriate

Accuracy alone is not a meaningful metric for this dataset.

---

Related Datasets

• A binary IDS version of this dataset is published separately for attack detection use cases.
• Both datasets share identical features and preprocessing logic.

---