title	PhishGuard Pro
emoji	🛡️
colorFrom	red
colorTo	gray
sdk	gradio
app_file	app.py
pinned	false
license	mit

🛡️ PhishGuard Pro: Advanced Hybrid AI Scam & Phishing Detector

▶️ Live Interactive Demo on Hugging Face Spaces

📌 Project Overview

PhishGuard Pro is an enterprise-grade AI tool designed to detect, analyze, and explain phishing emails, SMS scams (Smishing), and financial fraud attempts.

It leverages a Hybrid AI Architecture by combining fast, accurate Sequence Classification (BERT) with generative explainability (RAG + LLM) to not only flag malicious content but also provide actionable cybersecurity advice in English.

This project was developed as a robust portfolio piece demonstrating advanced AI Engineering skills in Fintech & Cybersecurity.

---

⚠️ Infrastructure & Performance Note (For Reviewers)

Demonstration Tier Restriction: This live demo is currently hosting a massive 7-Billion Parameter LLM (Zephyr-7B-beta) on a free CPU-only tier (2 vCPUs) for portfolio demonstration purposes.

While the BERT Classifier and Regex IoC engine will execute instantly, the final AI Incident Response generation may take 5-10 minutes to process due to the lack of hardware acceleration (GPU) on the free tier. In a production cloud environment with NVIDIA A100s, inference occurs in sub-seconds.

---

🛠️ What I Built vs. What is Ready-Made

To maintain transparency and highlight the specific engineering effort, here is a breakdown of my custom implementation versus the open-source tools utilized:

⚙️ What I Built (My Core AI Engineering Contribution)

1. Hybrid AI Pipeline Architecture: Architected a dual-stage inference pipeline, fusing fast BERT-based sequence classification with LLM-powered context reasoning to balance real-time latency with deep analytical accuracy.
2. Specialized Financial RAG Engineering: Curated and embedded a high-fidelity vector knowledge base focused on sophisticated attack vectors (e.g., Pig Butchering, CEO Fraud / BEC, advanced Smishing), enabling the AI to counter complex social engineering tactics.
3. Automated IoC Forensics Extraction: Engineered a deterministic Threat Intelligence layer utilizing Regex to parse raw inputs, instantly isolating Indicators of Compromise (malicious domains, burner emails, spoofed numbers) for immediate forensic visibility.
4. Guardrailed Prompt Design: Implemented strict, constraint-based prompt architectures that systematically mitigate LLM hallucination and force the generation of standardized, actionable Incident Response Plans.
5. Enterprise-Grade Analytical Dashboard: Developed a dynamic, responsive security terminal utilizing Gradio and Plotly to visually synthesize classification metrics, threat probabilities (Interactive Risk Gauge), and LLM reasoning into an intuitive analyst dashboard.

📦 Ready-Made Open-Source Models (The Giants I Stand On)

I integrated state-of-the-art free models to achieve maximum accuracy with zero deployment cost:

• Phishing Classifier: Auguzcht/securisense-phishing-detection (Fine-tuned BERT-base).
• Vector Embeddings: sentence-transformers/all-MiniLM-L6-v2 (Fast deployment embeddings).
• Reasoning Engine (LLM): HuggingFaceH4/zephyr-7b-beta (Highly capable instruction-tuned 7B model).
• Orchestration: LangChain (Vector DB bridging) and FAISS (In-memory similarity search).

---

🚀 How to Run Locally

1. Clone the repository and install dependencies:

   pip install -r requirements.txt

2. Run the Gradio app:

   python app.py

Note: Due to the usage of powerful LLMs, this application may require significant memory (RAM/VRAM) to run optimally locally. On Hugging Face Spaces, it runs efficiently within available hardware limits.

---

🔮 Future Roadmap (Enterprise Scaling)

While the current architecture serves as a highly effective Minimum Viable Product (MVP), transitioning this to a production-grade enterprise deployment would involve the following architectural upgrades:

1. Model Fine-Tuning (Data-Centric Optimization)

   • What: Fine-tuning the base Sequence Classification model (e.g., utilizing larger BERT variants) specifically on datasets containing high-volume contextual Smishing (SMS phishing) and WhatsApp fraud.
   • Why: Scammers rely heavily on social engineering specifically formatted for mobile platforms. Fine-tuning guarantees extreme precision against zero-day telecom fraud.

2. Live Threat Intelligence Integration (Dynamic Vector DB)

   • What: Migrating the static in-memory RAG store to a live, distributed Vector Database (such as Pinecone or Milvus) connected to automated OSINT (Open-Source Intelligence) threat feeds.
   • Why: Scam narratives evolve daily. A dynamic Vector DB ensures the AI's contextual knowledge base updates in real-time without requiring application rebuilds or downtime.

3. Active URL Sandboxing & API Verification

   • What: Automatically routing the extracted Indicators of Compromise (IoCs) through professional threat aggregation APIs like VirusTotal or Google Safe Browsing.
   • Why: While the current system excels at behavioral linguistic analysis, combining AI heuristics with deterministic IP/URL reputation checks provides a fail-proof, multi-layered security blanket.

4. Autonomous AI Agents (Tool-Calling Integration)

   • What: Upgrading the passive RAG pipeline into an active Autonomous Agent framework (via LangChain Agents) equipped with tools such as a SandboxBrowserTool and DomainLookupTool.
   • Why: Instead of merely analyzing the passive text of a message, an Agent can investigate it. If an email contains a link, the Agent autonomously securely browses the link, observes the webpage (e.g., detecting a cloned PayPal login screen), checks the domain registration date, and synthesizes a conclusive forensic report. This active investigation represents the true State-of-the-Art in AI Cybersecurity.

---

📜 Legal Disclaimer

This tool is for educational and advisory purposes only. Complex fraud schemes evolve rapidly. Always rely on authorized bank or official channels for final verification.