AntiFish

AntiFish detects, assists users with navigating, and spreads awareness of phishing and scam attempts on Email and SMS. Leveraged by ML models trained on real data and an AI assistant.

Comment

Inspiration

AntiFish started from personal stakes. One of my former professors was scammed out of $10,000, and it was a blunt reminder that these attacks target everyone. At the same time, my grandparents are constantly worried about whether a text, email, or phone call is legitimate, and they often don’t have a single place to check something quickly and confidently.

We wanted one place that combines detection + live awareness + education instead of treating them as separate tools. Phishing causes billions in losses each year precisely because awareness hasn’t reached everyone. AntiFish aims to close that gap with an end-to-end experience: from detecting a threat to understanding it to learning how to avoid the next one.

What it does

AntiFish is a phishing detection and education web app with three core parts:

  • Scam Risk Checker: paste a message (email/SMS) or describe a call and get a risk score with clear reasons (why it looks suspicious, what signals triggered).
  • Scam Tracker Dashboard: a “Downdetector-like” feed of trending scam campaigns so users can see what’s circulating right now and detect how similar other scams are to theirs.
  • Awareness Hub: curated phishing news + practical resources so users learn patterns and safer habits over time.

How we built it

  • ML risk scoring: Trained a TF-IDF + Logistic Regression classifier on public scam/phishing datasets to produce a baseline phishing probability, then blended it (60/40) with a lightweight rule-based overlay — flagging signals like urgency language, gift card requests, credential prompts, and suspicious links — into a final risk band (LOW / SUSPICIOUS / HIGH)
  • Web app + API: a front-end UI that calls a backend API for scoring, storing submissions, and powering the tracker feed.
  • Awareness content: structured “news items” and evergreen “resources” so the app isn’t just a checker; it’s a continuous learning surface.

Challenges

  • Balancing false positives vs. false negatives: we tuned thresholds so obvious scams flag hard, while still giving safe guidance for gray-area messages.
  • Getting explainability right: we focused on human-readable reasons instead of just a score.

What’s next

  • Add screenshot/URL analysis, better link reputation checks, and richer “next steps” flows (bank/identity protection checklists).
  • Improve model robustness with more diverse datasets and better calibration.
  • Expand the tracker with automated clustering of similar scam templates.
Share this project:

Updates