Hugnd-UIT/ML-Based-Malicious-Package-Detection

DySec

A Machine Learning-Based Dynamic Analysis For Detecting Malicious Packages In PyPI Ecosystem

Overview

Malicious python packages make software supply chains vulnerable by exploiting trust in open-source repositories like PyPI.
Lack of real-time behavioral monitoring renders metadata inspection and static code analysis inadequate against advanced attack strategies such as typosquatting, covert remote access activation, and dynamic payload generation.

To address these challenges, we introduce DySec, a machine-learning-based dynamic analysis framework for PyPI that uses eBPF kernel and user-level probes to monitor package behavior during installation.

DySec captures 36 real-time behavioral features, including:

• System calls
• Network traffic
• Resource usage
• Directory access
• Installation patterns

By analyzing these behaviors, DySec detects typosquatting, covert remote access activation, dynamic payload generation, and multiphase attack malware.

We developed a comprehensive dataset of 14,271 Python packages including 7,127 malicious samples executed in an isolated environment.

Experimental results demonstrate that DySec achieves 96% detection accuracy with <0.5 s inference latency after feature extraction, reducing false negatives by 78.65% compared to static analysis and 82.24% compared to metadata analysis.

During evaluation, DySec flagged eleven packages that PyPI classified as benign — manual analysis confirmed six as malicious, leading to four removals by PyPI maintainers.

DySec thus bridges the gap between reactive traditional methods and proactive, scalable threat mitigation in open-source ecosystems.

Figure 1: The overall workflow of DySec for detecting malicious packages.

---

Citation

If you use this study in your research, please cite it as:

@article{Mehedi2025DySec,
  author         = {Sk Tanzir Mehedi and Chadni Islam and Gowri Ramachandran and Raja Jurdak},
  title          = {DySec: A Machine Learning-based Dynamic Analysis for Detecting Malicious Packages in the PyPI Ecosystem},
  journal        = {arXiv preprint},
  year           = {2025},
  eprint         = {2503.00324},
  archive-prefix = {arXiv},
  primary-class  = {cs.CR},
  doi            = {10.48550/arXiv.2503.00324},
  url            = {https://arxiv.org/abs/2503.00324}
}

Authors

• Sk Tanzir Mehedi (Queensland University of Technology)
  ORCID
• Chadni Islam (Edith Cowan University)
  ORCID
• Gowri Ramachandran (Queensland University of Technology)
  ORCID
• Raja Jurdak (Queensland University of Technology)
  ORCID

Dataset

https://huggingface.co/datasets/Hugnd-UIT/ML-Based-Malicious-Package-Detection

Directory

Repository
│
├── data/
│   │
│   ├── select-data.py
│   └── split-data.py
│
├── dataset/
│   │
│   ├── QUT-DV25-Processed/
│   │   │
│   │   ├── QUT-DV25-Filetop-Traces/
│   │   ├── QUT-DV25-Install-Traces/
│   │   ├── QUT-DV25-Opensnoop-Traces/
│   │   ├── QUT-DV25-Pattern-Traces/
│   │   ├── QUT-DV25-SysCall-Traces/
│   │   └── QUT-DV25-TCP-Traces/
│   │
│   ├── QUT-DV25-Raw/
│   │   │
│   │   ├── QUT-DV25-Benign/
│   │   │   │
│   │   │   ├── QUT-DV25-Filetop-Traces/
│   │   │   ├── QUT-DV25-Installation-Traces/
│   │   │   ├── QUT-DV25-Opensnoop-Traces/
│   │   │   ├── QUT-DV25-PIDs/
│   │   │   ├── QUT-DV25-Pattern-Traces/
│   │   │   └── QUT-DV25-TCP-Traces/
│   │   │
│   │   └── QUT-DV25-Malicious/
│   │       │
│   │       ├── QUT-DV25-Filetop-Traces/
│   │       ├── QUT-DV25-Installation-Traces/
│   │       ├── QUT-DV25-Opensnoop-Traces/
│   │       ├── QUT-DV25-PIDs/
│   │       ├── QUT-DV25-Pattern-Traces/
│   │       └── QUT-DV25-TCP-Traces/
│   │
│   ├── QUT-DV25.csv
│   ├── README.md
│   ├── test.csv
│   ├── train.csv
│   └── val.csv
│
├── models/
│   │
│   ├── Decision-Tree.pkl
│   ├── Gradient-Boosting.pkl
│   ├── KNN.pkl
│   ├── Logistic-Regression.pkl
│   ├── Random-Forest.pkl
│   └── SVM.pkl
│
├── reports/
│   │
│   ├── confusion-matrices.png
│   ├── model-performance.csv
│   └── roc-curves.png
│
├── src/
│   │
│   ├── scripts/
│   │   │
│   │   ├── monitor.sh
│   │   ├── prerequisites.sh
│   │   └── trace.sh
│   │
│   ├── evaluate.py
│   ├── extract.py
│   ├── main.py
│   ├── predict.py
│   ├── preprocess.py
│   └── train.py
│
├── .gitignore
└── README.md

Usage

1. Setup

Install kernel monitoring dependencies, eBPF toolkits, strace, and python dependencies:

sudo bash src/scripts/prerequisites.sh
chmod +x src/scripts/monitor.sh
chmod +x src/scripts/trace.sh 

---

2. Pipeline

Run live dynamic tracing, feature extraction, and prediction:

sudo python3 src/main.py <package_name> [duration]

Example:

sudo python3 src/main.py requests 30

Skip re-installing:

python3 src/main.py requests --skip

3. Train & Evaluate model

python3 src/train.py

python3 src/evaluate.py

Report

1. Evaluation results

All 6 machine learning models were evaluated on the independent test dataset:

Model	Accuracy	Precision	Recall	F1-Score	ROC-AUC	FPR	FNR
Gradient Boosting	98.23%	98.22%	98.22%	98.22%	0.9986	1.77%	1.78%
Random Forest	96.36%	96.79%	95.88%	96.33%	0.9951	3.17%	4.12%
Decision Tree	95.00%	95.64%	94.29%	94.96%	0.9784	4.29%	5.71%
SVM	89.30%	90.86%	87.37%	89.08%	0.9557	8.77%	12.63%
Logistic Regression	86.74%	88.90%	83.91%	86.33%	0.9387	10.45%	16.09%
KNN	81.27%	82.49%	79.33%	80.88%	0.8891	16.79%	20.67%

Figure 2: Confusion matrices

Figure 3: Receiver Operating Characteristic (ROC) curves

---

2. Our results vs DySec results

Classifier	DySec Accuracy	Our Accuracy	Delta	Consistency
Random Forest	~95.99%	96.36%	+0.37%	Perfectly consistent with original paper
Gradient Boosting	~96.50% - 97.20%	98.23%	+1.03%	Superior generalization with lowest error rates
Decision Tree	~93.80% - 94.50%	95.00%	+0.50%	Consistent with tree baseline performance
SVM	~88.00% - 90.00%	89.30%	-0.20%	Highly aligned with linear boundary baselines

Analysis of discrepancies:

1. Partitioning random seeds: The published paper did not disclose the fixed pseudorandom seed used to partition QUT-DV25 into train, validation, and test splits. Our stratified split ensures a balanced distribution 1,072 Benign vs. 1,069 Malicious in test, stabilizing boundary estimations.
2. Hyperparameter tuning: Gradient Boosting achieved 98.23% accuracy due to hyperparameter calibration learning_rate=0.1, max_depth=5, n_estimators=100, effectively mitigating tree overfitting on sparse behavioral features.
3. Feature preprocessing & scaling: Implementing MinMaxScaler pipelines for distance/linear models SVM, Logistic Regression, KNN stabilized convergence across heterogeneous scale distributions of syscall and network counts.

---

3. Strengths & Limitations

A. Challenges:

• Undocumented trace parsing edge-cases: The original paper does not publish the exact regex parsers for edge-case syscall formats in strace or kernel timestamp mismatches between eBPF tools.

• Absence of standard split manifest: Neither specific package lists per split nor random seeds were provided in the public paper release, requiring an empirical stratified partition to ensure statistical reproducibility.

B. Limitations:

• Semantic loss in 36 SEF: By aggregating dynamic logs into 36 numerical counts, vital qualitative context is lost. Example distinguishing between a legitimate DNS resolution to pypi.org vs. a C2 reverse shell connection to an arbitrary raw IP 185.x.x.x:4444.

• Install-time execution confinement: Telemetry collection is restricted to the package installation phase (pip install). "Logic bombs" or evasion-aware malware that defer malicious execution until specific functions are imported at runtime bypass this install-time capture window.

C. Strengths:

• Immunity to Code Obfuscation: Traditional static analysis tools fail against base64 encoding, hex packing, and eval(exec()) obfuscation. In contrast, kernel-level eBPF probes inspect the executed behavior in the OS kernel regardless of how obfuscated the user-space Python script is.

• Ultra-low Latency Inference: Feature extraction and model prediction execute in <0.1 seconds, enabling real-time gating in CI/CD software supply chain pipelines.

License

This project is licensed under the MIT License - see the LICENSE file for details.