AI Data Pipeline Security Audit

A comprehensive security audit framework for multi-stage NLP data annotation pipelines. This project identifies, models, and mitigates attack surfaces in AI/ML data workflows — including prompt injection, data poisoning, model extraction, and data drift vectors.

Built as part of cybersecurity coursework at Pulchowk Campus, Institute of Engineering, with practical inspiration from real-world AI annotation workflows.

Features

• STRIDE-based threat modeling per pipeline stage
• Risk register with quantitative severity scoring (likelihood × impact)
• Statistical anomaly detection on annotation batches
• ISO 27001 Annex A control mapping
• Prompt injection pattern scanner
• Data poisoning / label flip detector
• Markdown + JSON audit report generation

Project Structure

ai-pipeline-security-audit/
├── src/
│   ├── threat_modeler.py
│   ├── risk_register.py
│   ├── anomaly_detector.py
│   ├── control_mapper.py
│   ├── prompt_injection_scanner.py
│   ├── data_poisoning_detector.py
│   ├── report_generator.py
│   └── utils.py
├── tests/
├── data/samples/
├── docs/
├── configs/pipeline_config.yaml
└── requirements.txt

Installation

git clone https://github.com/divaspoudel/ai-pipeline-security-audit.git
cd ai-pipeline-security-audit
pip install -r requirements.txt

Usage

# Full audit
python src/report_generator.py --config configs/pipeline_config.yaml --output reports/audit_report.md

# Scan for prompt injection
python src/prompt_injection_scanner.py --input data/samples/sample_annotations.json

# Detect data poisoning
python src/data_poisoning_detector.py --input data/samples/sample_annotations.json --threshold 0.15

License

MIT