silviocretti/DOS201X_highly_unbalanced

DoS201X Highly Unbalanced (Federated, Pre-Partitioned)

This dataset has been specifically prepared for federated learning (FL) experiments on DDoS attack detection. It is partitioned among 15 clients, with each client assigned a single attack type together with a portion of benign traffic, resulting in a non-IID distribution of network flows across clients. This partitioning is designed to reproduce a realistic FL setting in which individual clients have access to different attack types while sharing benign traffic. The dataset is a preprocessed, repartitioned derivative of the CIC-DDoS2019, CIC-IDS2018 and CIC-IDS2017 datasets, originally published by the Canadian Institute for Cybersecurity (CIC), University of New Brunswick.

More details about the CIC-DDoS2019 dataset can be found on this page and in the following scientific paper:

Iman Sharafaldin, Arash Habibi Lashkari, Saqib Hakak, and Ali A. Ghorbani, "Developing Realistic Distributed Denial of Service (DDoS) Attack Dataset and Taxonomy", IEEE 53rd International Carnahan Conference on Security Technology, Chennai, India, 2019.

More details about the CIC-IDS2018 dataset can be found on this page and in the following scientific paper:

Iman Sharafaldin, Arash Habibi Lashkari, and Ali A. Ghorbani, “Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Characterization”, 4th International Conference on Information Systems Security and Privacy (ICISSP), Portugal, January 2018

More details about the CIC-IDS2017 dataset can be found on this page and in the following scientific paper:

Iman Sharafaldin, Arash Habibi Lashkari, and Ali A. Ghorbani, “Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Characterization”, 4th International Conference on Information Systems Security and Privacy (ICISSP), Portugal, January 2018.

This version contains network flows partitioned per attack-type/client_id, converted from HDF5 to Parquet, and consolidated into three splits (train, val, test).

Dataset structure

Each split (train.parquet - ~80% of the whole dataset, val.parquet - ~10%, test.parquet - ~10%) contains one row per bidirectional network flow sample, with the following columns:

Column	Type	Description
client_id	string	Attack-type / client partition the sample belongs to (e.g. "11-NetBIOS")
features	list<float32>	Flattened feature window; reshape to (10, 32) per sample to restore the original layout
label	int64	0 = benign, 1 = malicious

Samples are grouped by client_id, mirroring the original per-client HDF5 directory structure, used in the federated learning experimentation. The original HDF5 structure comprises 15 clients X 3 splits = 45 HDF5 files in the form of arrays of shape n = 10 rows and f = 32 columns. See this repository for more details.

Each client contains samples of benign traffic and only one type of attack. Although each group has been balanced to ensure an approximately equal distribution between benign and DDoS samples, the partition across groups/clients is strongly non-i.i.d since each one represents a single attack type.

Related Work

This dataset has been used in:

Roberto Doriguzzi-Corin, Silvio Cretti, Silvio Ranise, Federated Learning in the Wild: A Comparative Study for Cybersecurity under Non-IID and Unbalanced Settings