Methodological Framework for Generating Synthetic, Balanced Datasets for Web Intrusion Detection using Automated Purple Teaming

The scarcity of high-quality, labeled datasets is a primary hindrance in the development of Deep Learning-based Web Application Firewalls (WAFs). Real-world traffic logs are often unavailable due to privacy regulations or lack ground-truth labels, leading to models that suffer from high false-positive rates. This technical report presents a comprehensive methodology for the generation of synthetic datasets through an automated "Purple Teaming" approach. By orchestrating a custom-developed offensive framework against a controlled ecosystem of vulnerable microservices, we establish a reproducible pipeline to capture, tokenize, and label HTTP traffic. This method ensures class balance between benign and malicious requests, covering the full spectrum of OWASP Top 10 threats, and provides a sustainable alternative to static, outdated public datasets.