lacg030175/CIC-IoT-2023-neto-full

CIC-IoT-2023 — Neto-Full (Authoritative 46.7M)

This is the authoritative canonical CIC-IoT-2023 dataset at the row count published by Neto et al. (2023): 46,686,579 rows × 46 features.

Sourced from the Kaggle mirror akashdogra/ciciot23csv (13.75 GB single CSV) which itself was derived from CIC's official 169-file distribution. Compared to bencorn's HF mirror (45M, 39 features), this preserves:

• —All 46 original features (bencorn dropped 7)
• —The full 46.7M row count (bencorn re-merge lost ~1.7M)

NaN/Inf preservation: NaN/±inf rows are kept (no dropna). Pair with ThermometerEncoder(invalid_encoding="single_bit") to encode invalid states as a learnable per-feature flag bit instead of silently collapsing to zero.

Splits

• —random_3way: 80% train / 10% test / 10% validation (stratified on binary label, seed=42)
• —random: 80% train / 20% test (test = test ∪ validation from random_3way)

Provenance

• —Original: Neto et al. (2023), "CICIoT2023: A real-time dataset and benchmark for large-scale attacks in IoT environment", Sensors 2023.
• —Distribution: CIC at unb.ca/cic/datasets/iotdataset-2023.html
• —This HF dataset preserves the original Label (canonicalized, mixed case) AND Label_orig (original UPPERCASE), plus normalized attack_class and binary label.

Class distribution

• —DDoS: 33,984,560 (72.79%)
• —DoS: 8,090,738 (17.33%)
• —Mirai: 2,634,124 (5.64%)
• —Benign: 1,098,195 (2.35%)
• —Spoofing: 486,504 (1.04%)
• —Recon: 354,565 (0.76%)
• —Web-based: 24,829 (0.05%)
• —BruteForce: 13,064 (0.03%)