Graduation Project 2025 — Web Application Security

Hybrid Web Application Firewall (Hybrid WAF)

Seven graduated filtering layers plus an LLM arbitration layer.

Year2025
RoleGraduation project — design, development, testing, documentation
StatusBeing prepared for public open-source release

Project Summary

A hybrid architecture that reconciles accuracy with performance through a cheapest-first principle. Seven graduated layers resolve clear cases cheaply, invoking large models only on ambiguous ones.

Web SecurityWAFAI ArbitrationSQLi/XSS/SSRF

The Problem

Rule-based firewalls fail against obfuscated payloads; raising sensitivity multiplies false positives until the site is broken for its users; and sending every request to an LLM is too slow and expensive for production.

The project addresses all three with one principle: lowest cost first — cheap layers settle the obvious cases, and large models are only invoked for grey areas.

Key Features

Seven layers

Protocol, reputation, rate-limit, signatures, semantic, behavioral, AI.

Gray-area AI

Models invoked only when needed.

Structured output

Allow/block, confidence, attack class.

Dual-model cross-check

Compared outputs for higher confidence.

Cumulative scoring

Aggregate risk signals across layers.

Feedback loop

Log arbitration decisions to improve rules.

Technical Challenges & Design Decisions

  • Judging grey areas: the first six layers either produce a clear decision or nothing. The seventh (LLM arbitration) is invoked only on uncertainty — which is what makes the system production-viable rather than a research demo.
  • Structured output is a prerequisite: you cannot automate a decision built on free text. Prompt engineering forces the model to return an analysable decision (allow/block, confidence, attack class), and the second model acts as a cross-check on critical cases.
  • Camouflaged payloads: surface inspection is fooled by repeated encoding. Payloads are therefore decoded in multiple stages and normalised before signature matching.
  • Continuous feedback: every arbitration decision is logged so it can become a static rule later, moving the system from "a model called often" towards "a model called rarely".

Technical Architecture

Seven-layer pipeline — a request moves layer by layer, and LLM arbitration is reached only at the end
مخطط الطبقات السبع — الطلب يمرّ من طبقة إلى أخرى، ولا تُستدعى نماذج اللغة إلا في نهاية المسار
FastAPIReverse-proxy middleware.
RedisRate-limit counters.
SQLiteEvent log.
DockerReproducible packaging.
BashOps scripts.

Testing Methodology

  • OWASP Top 10 payload classes including obfuscated SQLi, XSS and SSRF, to verify layer-4 signatures still fire after normalisation.
  • Measured false positives before and after enabling the AI arbitration layer.
  • Measured response latency and the share of requests that actually reach arbitration — the metric that proves the lowest-cost-first principle works.
  • Re-tested attack scenarios after every change to the layer map.

Outcome & Impact

  • Obvious cases settled by cheap layers, so the normal path sees no noticeable latency.
  • Large-model calls are bounded to grey-area cases rather than every request.
  • A feedback loop: each grey case becomes a rule, so reliance on inference falls over time.

Full detail of the seven layers and the methodology is in the case study. Status: being prepared for public open-source release.

Read the full case study Being prepared for public release