---
license: mit
---

# SMU\_MalwareDetection PE Assembly Dataset

This dataset consists of assembly code fragments extracted from malicious and benign Portable Executable (PE) files. It was created to support deep learning-based malware classification and **concept drift experiments** across malware families and time periods. Assembly code was segmented using a sliding window technique to preserve contextual information.

## ⚠️ Note: Class Imbalance

This dataset has a **class imbalance**, with fewer benign samples (`0` label) compared to malware samples (`1` label).
While the difference is not extreme, users should be aware of this imbalance and consider applying appropriate techniques (e.g., class weighting, sampling strategies) during training or evaluation.

## 🦠 Dataset Composition (Malware Families)

The dataset is intentionally composed of samples from **three specific malware families** to support concept drift studies:

* **RemcosRAT** (20 files)
* **AgentTesla** (45 files)
* **GuLoader** (20 files)

Each malware family includes samples from various years to allow for temporal drift analysis.

In total:

* **85 malware PE files**
* **85 benign PE files**

## Dataset Description

The dataset includes assembly code segments derived from two types of PE files:

* **Malware Samples**
  Malicious PE files were downloaded from [MalwareBazaar](https://bazaar.abuse.ch/), with a focus on samples uploaded in **multiple years**. Disassembly was performed using `objdump` on Ubuntu.

* **Benign Samples**
  Benign PE files were collected from trusted executable files found on a local machine using PowerShell, then disassembled similarly using `objdump`.

## Sliding Window Preprocessing

To prepare the data for machine learning:

* Assembly code was segmented using a **sliding window approach**:

  * **Window Size:** 500 lines
  * **Stride:** 175 lines (overlap between windows)
* This method preserves contextual relationships between instructions.
* All segments generated from a **single file are grouped together** to avoid leakage between training and test sets.
* **Raw assembly code** is used as-is; no tokenization or preprocessing beyond segmentation was applied.

## Data Format

Each entry in the dataset includes:

* `assembly`: A segment of assembly code extracted via the sliding window
* `label`: A binary classification label

  * `0` = benign
  * `1` = malware

## Intended Use

This dataset is suitable for the following tasks:

* Malware classification using NLP-based models (e.g., Transformers)
* Research on **concept drift in malware detection**
* Context-aware malware analysis using segmented disassembly

## Tools Used

* `objdump` (for disassembly)
* PowerShell (for benign file collection)
* Python (for segmentation and grouping)