💻 computer science

QUT-DV25: A Dataset for Dynamic Analysis of Next-Gen Software Supply Chain Attacks

The paper introduces QUT-DV25, a new dynamic analysis dataset featuring real-time execution traces from over 14,000 Python packages, designed to bridge the gap in detecting sophisticated, multi-phase software supply chain attacks that evade traditional static and metadata-based detection methods.

Original authors: Sk Tanzir Mehedi, Raja Jurdak, Chadni Islam, Gowri Ramachandran

Published 2026-02-12

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Sk Tanzir Mehedi, Raja Jurdak, Chadni Islam, Gowri Ramachandran

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The "Trojan Horse" in Your Digital Toolbox: Explaining QUT-DV25

Imagine you are a professional carpenter. To do your job, you buy tools from a massive, global warehouse called "The Tool Depot" (this is like PyPI, the place where programmers get pre-made code). Most tools are great, but occasionally, a bad actor sneaks a "Trojan Horse" into the warehouse.

These aren't just broken hammers. These are "smart" tools. A screwdriver might look perfectly normal on the shelf, but the moment you bring it into your workshop and start using it, a tiny hidden compartment opens up, a miniature spy crawls out, and starts stealing your blueprints or unlocking your back door.

The Problem: The "Photo-Only" Security Guard

Currently, the "security guards" protecting the digital warehouse mostly look at the labels on the boxes (Metadata) or the blueprints of the tools (Static Analysis).

If a thief is clever, they can make a box look identical to a famous brand and make the blueprints look perfectly safe. The security guard sees a "Standard Hammer" label and a "Solid Steel" blueprint and lets it through. The guard never actually uses the hammer to see if it secretly contains a spy. This is why modern "next-gen" attacks are so successful—they wait until they are inside your house before they reveal their true nature.

The Solution: QUT-DV25 (The "High-Tech Observation Lab")

A team of researchers from Australia has built something much smarter: QUT-DV25.

Instead of just looking at the box, they built a high-tech, isolated "Testing Lab" (a Sandbox). When a new tool arrives, they don't just look at it; they actually take it into the lab, put it to work, and watch it through a microscopic, high-speed camera.

To do this, they use a technology called eBPF. Think of eBPF as a "super-sensor" embedded in the very walls of the lab. It doesn't just watch the tool; it watches every single tiny movement:

The "Footprints" (File I/O): Did the tool move files it wasn't supposed to?
The "Phone Calls" (Network Traffic): Did the tool try to "call home" to a suspicious stranger?
The "Heartbeat" (System Calls): Did the tool suddenly start acting erratic or trying to take control of the lab's power supply?

What did they find?

The researchers tested over 14,000 "tools" and found that a huge number of them were actually hiding malicious secrets.

The most impressive part? Their "super-sensor" system was so good that it caught four malicious packages that everyone else thought were perfectly safe. These packages had thousands of downloads and were sitting in the "warehouse" undetected. Once the researchers found them, they reported them, and the bad packages were removed.

Why does this matter to you?

Even if you aren't a programmer, you live in a world built on this code. The apps on your phone, the software in your car, and the systems running your bank all rely on these "tools."

By creating QUT-DV25, these researchers have provided a new "training manual" for AI security guards. They have given the world a way to move past "looking at the label" and start "watching the behavior," making it much harder for digital spies to hide in plain sight.

Technical Summary: QUT-DV25: A Dataset for Dynamic Analysis of Next-Gen Software Supply Chain Attacks

1. Problem Statement

The rapid growth of the Python Package Index (PyPI) has made it a primary target for sophisticated software supply chain attacks. Modern adversaries have moved beyond simple malicious code to "next-gen" multi-stage attacks, including:

Typosquatting: Mimicking legitimate package names.
Dynamic Payload Generation: Executing malicious logic only under specific environmental conditions.
Multi-phase Execution: Separating the installation phase from the actual malicious activity (e.g., remote access activation or data exfiltration) to evade detection.

Existing datasets used to train Malicious Detection Systems (MDS) are largely limited to metadata inspection (author profiles, descriptions) or static code analysis (function signatures, imports). These methods fail to capture the behavioral "footprint" of a package during and after installation, leaving a critical gap in detecting threats that remain dormant until runtime.

2. Methodology

The authors propose a novel framework for constructing a dynamic analysis dataset through a controlled, isolated testbed.

Testbed Configuration: The researchers utilized a cluster of 16 Raspberry Pi devices running Ubuntu, isolated within a private network to prevent accidental malware propagation. Monitoring is implemented via eBPF (extended Berkeley Packet Filter), which allows for low-overhead, real-time tracing of kernel and user-space events without modifying the Linux kernel.
Dataset Construction (QUT-DV25):
- Collection: Malicious packages were gathered from threat intelligence sources (e.g., GitHub advisories). Benign counterparts were identified using lexical similarity algorithms to ensure a balanced and comparable dataset.
- Validation: Labels were verified using multiple validators (VirusTotal, NDV, Snyk) and manual inspection.
- Trace Extraction: Packages were installed in the sandbox. The system captured traces during the install-time and a fixed 120-second post-install-time window (determined empirically to be sufficient to capture most behavioral divergences).
Feature Engineering: The authors extracted 36 real-time features across six categories:
1. File I/O: Abnormal file access/creation.
2. Dependency Logs: Indirect malicious installations.
3. File Open Attempts: Access to protected directories.
4. TCP Traffic: Connections to suspicious remote endpoints.
5. System Calls: Privilege misuse or sabotage.
6. Behavioral Patterns: Sequences of events (e.g., fork $\rightarrow$ execve).

3. Key Contributions

The QUT-DV25 Dataset: A large-scale dynamic dataset comprising 14,271 Python packages (7,127 malicious and 7,144 benign).
eBPF-based Monitoring Framework: A scalable methodology for capturing granular, real-time system-level behaviors in an isolated environment.
Baseline Benchmarks: A comprehensive evaluation of four Machine Learning (ML) models (Random Forest, Decision Tree, SVM, Gradient Boosting) and two Deep Learning (DL) models (CNN, Transformer) on the new dataset.

4. Results

Model Performance: The Random Forest (RF) model using the "Combined Traces" feature set achieved the highest performance, with an accuracy of 95.99% and an F1-score of 96.02%.
Superiority over Baselines: QUT-DV25 significantly outperformed existing metadata-only and static-only datasets. While metadata datasets suffered from high false positives (13.90% FPR), and static datasets had higher false negatives (6.63% FNR), QUT-DV25 provided a more robust balance.
Real-world Impact: Using the dataset to train ML models, the researchers identified four previously undetected malicious PyPI packages that had thousands of downloads. These packages were reported to PyPI maintainers and subsequently removed from the ecosystem.
Efficiency: While DL models (CNN/Transformer) achieved comparable accuracy to ML models, the ML models (specifically RF and DT) were significantly faster in terms of training and testing time, making them more suitable for real-time deployment.

5. Significance

This work bridges the gap between static and dynamic analysis in software supply chain security. By providing a dataset that captures install-time and post-install-time behaviors, QUT-DV25 enables the development of more resilient Malicious Detection Systems capable of identifying multi-stage, polymorphic, and environment-triggered attacks. The practical validation—resulting in the removal of live malicious packages—demonstrates that dynamic behavioral analysis is a vital component of modern cybersecurity defense for open-source ecosystems.