README.md

---
pretty_name: "MalayaNetwork_GT: Non-IID dataset for FL-Based network traffic classification evaluation"
dataset_name: MalayaNetwork_GT
tags:
  - network-traffic
  - cybersecurity
  - network-traffic-classification
  - non-iid
  - federated-learning
task_categories:
  - tabular-classification
license: cc-by-4.0
configs:
  - config_name: by_application
    data_files:
      - split: bittorent
        path: "csv_output/Bittorent/*.csv"
      - split: chrome_rdp
        path: "csv_output/ChromeRDP/*.csv"
      - split: discord
        path: "csv_output/Discord/*.csv"
      - split: ea_origin
        path: "csv_output/EA Origin/*.csv"
      - split: microsoft_teams
        path: "csv_output/Microsoft Teams/*.csv"
      - split: slack
        path: "csv_output/Slack/*.csv"
      - split: steam
        path: "csv_output/Steam/*.csv"
      - split: teamviewer
        path: "csv_output/Teamviewer/*.csv"
      - split: webex
        path: "csv_output/Webex/*.csv"
      - split: zoom
        path: "csv_output/Zoom/*.csv"
---

## Dataset Summary

**MalayaNetwork_GT** is a private network traffic dataset captured from several real network environments and multiple time windows to intentionally introduce **non-IID** characteristics. This repository contains:

- **Raw packet captures** (`PCAP/`): `.pcapng` files (approximately **35GB** total in the full dataset).
- **Derived flow features** (`csv_output/`): flow-based CSVs extracted using [CICFlowMeter](https://github.com/hieulw/cicflowmeter/tree/master) (CSV size **varies** depending on capture duration and traffic volume).

In this repo snapshot, there are **10 application/service classes** (one folder per class).

## Motivation

Many public network-traffic datasets (e.g., ISCX-VPN2016) can be **outdated** and may not reflect the **non-IID** nature of traffic seen in real deployments. For network traffic, *non-IID* commonly means the data distribution is not uniform across classes, environments, or time. For example, some traffic patterns appear only in certain environments or times of day.

This matters especially for **distributed learning** settings such as **Federated Learning (FL)**, where statistical heterogeneity can degrade convergence speed and model accuracy if clients observe substantially different traffic distributions.

## Dataset Creation

### Data collection

Traffic was captured using **Wireshark** from a network interface with **promiscuous mode** enabled to capture all frames (useful especially on wireless networks).

To encourage non-IID characteristics, captures were performed over a **three-month period** at different times of day:

- **Morning**: 9-10 AM
- **Peak hours**: 12-1 PM
- **Non-peak/night**: 8-9 PM

This introduces **temporal heterogeneity** (and potential concept drift): the same label may exhibit different feature distributions depending on time-of-day usage patterns.

Additionally, traffic was collected from **several network environments**, contributing to **covariate shift** between subsets (e.g., different user behavior and usage patterns across environments such as campus networks vs public Wi‑Fi vs home broadband vs mobile networks).

To minimize background traffic and support anonymity, we used the following methods:

- Dedicated **virtual machines** per application/service.
- Pseudo user accounts for registration where needed.
- Background noise reduction by removing bloatware, disabling automatic updates, and stopping non-essential processes.

Most services used **TLS-encrypted** payloads. **BitTorrent** may include unencrypted peer communications depending on peers and client behavior.

### Feature extraction

Packet captures (`.pcapng/.pcap`) were transformed into **flow-based features** using [CICFlowMeter](https://github.com/hieulw/cicflowmeter/tree/master) by Canadian Institute for Cybersecurity (CIC). The resulting CSVs include per-flow fields such as IP/ports/protocol, timestamps, flow duration, byte/packet rates, IAT statistics, packet length statistics, TCP flag counts, active/idle stats, and subflow aggregates. Most of these services encrypted their communication 
payload via TLS, with the exception of BitTorrent. Although BitTorrent clients such as 
Q-BitTorrent enabled encryption by default, not all P2P peers supported encrypted 
communication.


## Dataset Structure

### Repository layout

- **`PCAP/`**: raw captures grouped by class (subfolders contain `.pcapng/.pcap`)
  - Example: `PCAP/Zoom/Zoom_5G_Daytime1.pcapng`
- **`csv_output/`**: flow-feature CSV outputs (mirrors the class folder structure)
- **`cicflowmeter_batch.sh`**: batch runner to process each immediate subfolder under `PCAP/`. Used for reproducing the the csv outputs.

### Application/Service classes in this repo

- `Bittorent`
- `ChromeRDP`
- `Discord`
- `EA Origin`
- `Microsoft Teams`
- `Slack`
- `Steam`
- `Teamviewer`
- `Webex`
- `Zoom`

### Example file naming convention

Files commonly encode **environment** and **time window**, e.g.:

- `Zoom_5G_Daytime1.pcapng` / `Zoom_5G_Daytime1.csv`
- `Steam_Campus_Daytime1.pcapng` / `Steam_Campus_Daytime1.csv`
- `Webex_HomeBroadband_Daytime1.pcapng` / `Webex_HomeBroadband_Daytime1.csv`

## Intended Uses

- **Network Traffic classification** across application/service classes using flow features.
- **Non-IID / FL evaluation**

## Limitations

- **Non-IID by design**: results may differ significantly depending on how you split data (by environment, time, capture session).
- **Dataset size**:
  - PCAPs are ~**35GB** for the full dataset (private); CSV size depends on extracted flows and can vary widely.
- **Privacy**: raw captures and derived flows may include IP addresses and ports. This repo does not claim anonymization; treat the data as potentially sensitive.

## Licensing

This dataset is licensed under **Creative Commons Attribution 4.0 International (CC BY 4.0)**. See [`LICENSE`](LICENSE).

## How to reproduce the CSVs (CICFlowMeter)

### Prerequisites

- **Python**: 3.12+
- **CICFlowMeter**: via Python dependency `cicflowmeter`
- Installs may require system **libpcap** and **tcpdump** packages via apt/pacman/etc.

### Install

This repo includes `pyproject.toml`, so you can install dependencies with either `uv` or `pip`.

- Using `uv`:

```bash
uv sync
```

- Using `pip`:

```bash
pip install -r requirements.txt
```

### Generate CSVs from all folders in `PCAP/` using the batch script

Run CICFlowMeter once per subfolder under `PCAP/` and write outputs to `csv_output/<folder>/`:

```bash
./cicflowmeter_batch.sh -i ./PCAP -o ./csv_output
```

If you prefer a single combined CSV output directory:

```bash
./cicflowmeter_batch.sh -i ./PCAP -o ./csv_output --flat
```

### Generate CSVs from a single folder using CICFlowMeter directly

```bash
cicflowmeter -d "./PCAP/Zoom/" -c "./csv_output/Zoom/"
```

## Citation

If you use this dataset, please cite the associated thesis/work:

```bibtex
@misc{MalayaNetwork_GT,
  title     = {MalayaNetwork_GT: Non-IID dataset for FL-Based network traffic classification evaluation},
  author    = {Ariffin, Azizi and Haris, Afif},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Afifhaziq/MalayaNetwork_GT}
}
```