PhishGuard AI+

• 

  the ui/ux of PhishGuard-AI

• 

  Tests our deterministic SSL audit layer.

• 

  ests your Shannon Entropy math. It detects that the domain was likely generated by a bot.

• 

  Tests your Brand Heuristics + Risky TLD (.tk).

• 

  Tests your Google DNS API integration.

• 

  Proves our system has a high UX priority and doesn't flag trusted giants.

• 

• 

• 

  a working extension is made

🛡️ Inspiration

Phishing is the #1 cyber threat globally, but traditional protection relies on "Blacklists" that are often outdated within hours. Scammers create "Zero-Day" sites that bypass these lists easily. I was inspired to build a tool that doesn't just check a list, but actually analyzes the "DNA" of a URL in real-time to predict threats before a user even clicks.

🧠 How I Built It (accuracy = 91.4%)

I started with an academic baseline from 2016 which achieved 86.4% accuracy. However, modern threats are more sophisticated. I modernized the system through a 3-stage engineering process:

1. Data Modernization: I scrapped the dead Alexa Rank API and integrated the Majestic Million dataset (10 lakh rows) to create a real-world trusted whitelist.
2. Feature Engineering: I implemented Shannon Entropy to detect bot-generated "gibberish" domains and coded custom brand-spoofing heuristics.
3. ML Optimization: Using XGBoost in Google Colab, I performed hyperparameter tuning (1,000 trees, depth of 12) to reach a final validation accuracy of 91.4%.

🚀 The Breakthrough: Two-Stage Nested Architecture

During development, I identified a "Feature Dominance" bottleneck: URL Length held 68% of the predictive weight, making the AI "lazy." It would block safe long URLs (Wikipedia) and miss dangerous short links (Bitly).

I acted as a Data Scientist to solve this by building a Hierarchical Nested Logic:

• Outer Gate: Uses length to set the context.
• Inner Brain: Applies "Conditional Sensitivity." It gives long URLs a "Fair Trial" while being strictly aggressive on obfuscated short-links.

🔍 Triple-Audit Layer

Beyond ML, I implemented deterministic security audits:

• Protocol Audit: Instant flagging of insecure http connections.
• Infrastructure Audit: Integration with the Google DNS JSON API to verify domain integrity in real-time.

🛠️ Challenges Faced

Overcoming False Positives was the biggest hurdle. High-entropy UUIDs in modern apps (like ChatGPT) initially triggered the AI. Solving this required a balance between my high-sensitivity inference engine and a robust whitelisting system, teaching me the value of "Security with Usability."

📊 Risk Scoring Rubric

Built With

• javascript
• manifest-v3
• pandas
• python
• scikit-learn
• xgboost