UIA-IOT-IDS dataset: Comprehensive IoT network intrusion detection and sensor-anomaly
Description
This record contains two complementary, fully labelled datasets for machine-learning-based IoT security, together with the complete acquisition and analysis code. All data were collected from a single physical smart-home testbed at the University of Agder (Grimstad, Norway), built around a resource-constrained Raspberry Pi 3 gateway, five ESP8266 + DHT22 temperature/humidity nodes in distinct rooms, a PySense board, a Flask IoT server with InfluxDB storage, and three Kali Linux attacker VMs. Network flows were computed on-device in real time by a custom tcpdump-based extractor, so no raw PCAP is stored. The deposit is organised into two datasets that serve two tasks: (1) IDS_Dataset - network-flow INTRUSION DETECTION. The "Original" folder holds the raw, as-collected flows: 3,637,408 labelled bidirectional records across 54 classes, each described by 75 columns (6 identifier/metadata fields + 69 statistical features) plus a label. Of the 69 features, 46 are schema-compatible with CIC-IoT-2023 and 23 are additional bidirectional descriptors (per-direction packet counts, packet-length and inter-arrival-time statistics, and flow rates). The "Balanced" folder holds a cleaned, SMOTE-balanced version: 1,415,074 records, one CSV per class, with identifier and zero-variance columns removed (64 features + label). The taxonomy spans 51 attack sub-types in nine categories - Reconnaissance, DoS, DDoS, Brute Force, Web-based, Mirai, Spoofing, False Data Injection and Backdoor/C2 - plus three benign traffic types (sensor, SSH, Grafana). It reproduces the eight CIC-IoT-2023 categories and adds families absent from common benchmarks: False Data Injection (direct injection and man-in-the-middle sensor replay) and Backdoor/command-and-control beaconing. (2) AD_Dataset - sensor ANOMALY DETECTION. The "Benign" folder holds timestamped temperature and humidity readings from the five DHT22 nodes (one CSV per node); the "Abnormal" folder holds readings from three anomaly sources: a heat-biased PySense board (hardware fault), direct false-data injection, and ARP-spoofing replay. 4,253,024 readings fall within the 1 March - 13 April 2026 window analysed in the article; earlier readings are also included. Each reading is stored as one row carrying either a temperature or a humidity value (InfluxDB export format). All tabular data are UTF-8 CSV files with a header row; every network-flow record carries an explicit label. The deposit also includes the full code: the real-time flow extractor, the Flask server with intentionally vulnerable endpoints, the attack and injection scripts, the SMOTE preprocessing, and the model code (XGBoost benchmark, LSTM-autoencoder, MAML, Prototypical Networks). A disjoint base/novel class split supports supervised, unsupervised and few-shot detection research.
Files
Steps to reproduce
The full testbed, capture-mode configurations and per-class analyses are reported in the accompanying Data in Brief article; the collection and processing pipeline is summarised below. 1. Testbed. A Raspberry Pi 3 (Raspbian 11) runs a Flask IoT server that stores sensor readings in InfluxDB and visualises them in Grafana. Five ESP8266 + DHT22 nodes post temperature/humidity from five rooms; a PySense board provides a hardware-biased sensor. Three Kali Linux VMs act as attackers. 2. Benign collection. With only legitimate traffic running, sensor readings are logged to InfluxDB and exported as the AD_Dataset "Benign" CSVs, while the flow extractor (network_extractor.py, normal mode) records benign network flows. 3. Attack and anomaly generation. Each attack sub-type is run in its own session with the extractor set to the matching capture mode (normal / dos / ddos / recon / backdoor / mitm), which fixes the packet-grouping key for that family. Reconnaissance uses Nmap, Masscan, Nikto and fping; flooding uses hping3; Brute Force uses Hydra; Web attacks use SQLmap and curl against the vulnerable endpoints; Backdoor/C2 traffic is produced by backdoor.py on the Pi. Sensor anomalies are produced by the PySense board, by direct false-data injection, and by Scapy/arpspoof replay, giving the AD_Dataset "Abnormal" CSVs. 4. Labelling. Each captured CSV is tagged with its attack sub-type or benign class, and the per-sub-type files are concatenated to form IDS_Dataset/Original. 5. Cleaning and balancing. smote_balance.py removes duplicate and all-zero rows, drops identifier and zero-variance columns, and applies a tiered SMOTE over-sampling / random under-sampling scheme (benign capped; attack classes raised or lowered to 10k/20k/30k targets). The result is IDS_Dataset/Balanced. 6. Technical validation (optional). benchmark_models.py reproduces the XGBoost flow-classification baselines (10-category and 54-class); lstm_anomaly_iot.py trains one LSTM-autoencoder per benign sensor and flags anomalies at the 95th-percentile reconstruction-error threshold; protonet_iot.py and maml_iot.py reproduce the few-shot detectors. Software: Python 3, tcpdump, Flask, InfluxDB, scikit-learn, imbalanced-learn, XGBoost, PyTorch. The file paths and device IPs defined at the top of each script must be set to the local environment before running.
Institutions
- University of AgderAgder, Kristiansand