Inspiration

Scams and phishing attacks have evolved far beyond suspicious emails. Today, bad actors use text messages, spoofed voice calls, and manipulated images to trick people — especially students and less tech-savvy users — into giving up sensitive information or money. We built AI Shield because most existing detection tools focus on just one channel (email, or SMS, or images) and miss the bigger picture. We wanted a single system that could catch scams no matter what form they take.

What it does

AI Shield is a multi-modal scam and phishing detection system that analyzes text, voice, and image inputs to flag potentially fraudulent or malicious content. It:

Scans text messages, emails, and links for phishing indicators
Analyzes voice/call transcripts for common scam call patterns
Inspects images (e.g. screenshots, fake login pages) for visual signs of spoofing or manipulation
Outputs a risk score/classification so users can quickly tell if something is safe or suspicious

How we built it

We built AI Shield around a traditional machine learning classification pipeline:

Collected and preprocessed labeled datasets across each modality (text, voice/audio-to-text, and image)
Engineered features specific to each input type (e.g. linguistic/keyword features for text, metadata and visual features for images)
Trained separate classifiers per modality, then combined their outputs into a unified risk decision
Built a simple interface so users can submit content and get an instant classification result

Challenges we ran into

Aligning data from very different modalities (text, audio, images) into a consistent pipeline
Balancing false positives vs. false negatives — being too aggressive flags legitimate messages, being too lenient misses real scams
Limited labeled training data for certain scam types, especially voice-based scams
Making the system fast enough to feel usable in real time

Accomplishments that we're proud of

Successfully built a working multi-modal pipeline that handles three different input types
Achieved solid classification accuracy across each detection module
Created a system that's genuinely usable, not just a proof of concept

What we learned

How to design and combine multiple classifiers into one cohesive decision system
The real-world difficulty of collecting quality labeled data for scam/phishing detection
The importance of interpretability — users trust a "why" more than just a score

What's next for AI Shield

Expand training data to cover more scam types and languages
Explore deep learning models to improve accuracy beyond the current classifiers
Add browser/app integration so AI Shield can flag threats in real time as users browse or receive messages