Inspiration
Phishing scams disproportionately target vulnerable populations. These attacks can devastate individuals who can't afford to recover from financial fraud. I wanted to create a free, accessible tool that not only detects phishing attempts but educates users about why an email is suspicious, empowering them to protect themselves.
What it does
Fisher is an intelligent phishing detector that combines machine learning with rule-based analysis to identify malicious emails. It provides: ML-powered classification across 12 email categories (legitimate vs. various scam types) Suspicious URL detection (shortened links, suspicious TLDs, known phishing domains) Typo analysis (phishing emails often contain spelling errors) Explainable results with confidence scores and detailed breakdowns
How we built it
Tech Stack: Python 3.11 scikit-learn (Random Forest Classifier) Streamlit (UI framework) autocorrect (typo detection) regex (URL pattern matching) The Architecture : Data Collection & Cleaning: Curated and labeled 360+ email examples across 12 categories (real bank, job, delivery messages + their phishing counterparts) ML Model Training: Random Forest classifier achieving 80% accuracy and 0.78 average confidence on test data Feature Engineering: Removed URLs before typo checking to avoid false positives; implemented threshold-based similarity scoring Integration Layer: Combined ML predictions, URL analysis, and typo detection into unified risk score UI Development: Clean Streamlit + CSS interface with real-time analysis and detailed reporting
Challenges we ran into
Technical: Limited Hardware: Developed entirely on a 4GB RAM laptop with a failing HDD. Training iterations took 2+ minutes each, requiring strategic batching and patience. Dependency Hell: Python package conflicts (especially with spell-checking libraries) consumed hours of debugging. Eventually settled on autocorrect after trying 4+ alternatives. Data Quality: The initial dataset had mislabeled examples causing confidence scores to plummet. I had to scrap 1,300 lines and rebuild with clean data, a painful but necessary lesson in "garbage in, garbage out." Model Selection: I tested Logistic Regression, Naive Bayes, and Random Forest. Random Forest performed best in practice despite lower theoretical accuracy scores. Personal: Academic Pressure: I built this while completing Cambridge A-level final exams in Mathematics, Computer Science, and Physics. The development happened in sprints between exam sessions. Resource Constraints: I had limited internet access which meant offline development and strategic use of cached documentation
Accomplishments that we're proud of
Being able to complete this in a short span of time. And being able to participate in this.
What we learned
ML Engineering is not Just Algorithms: Data cleaning and quality control are 80% of the work. A simpler model on clean data outperforms a complex model on messy data. User-Centric Design: Initially focused on "correcting" typos, but users benefit more from seeing what was flagged. Resilience Through Constraints: Limited hardware forced me to write efficient code Python Ecosystem Challenges: Dependency management and package compatibility are real engineering problems Increase in the Sophistication of Scams: I had to refer to outside sources for help labeling my datasets because even I could not tell whether they were real or scam messages.
What's next for Phisher
Interactive Educational Game : Practice identifying scams of different sophistication to be able to spot them easily .
Built With
- css
- natural-language-processing
- python
- regex
- scikitlearn
- streamlit