October 2026
Science Fair: Vishing Detector
My 2026-27 science fair project: catching AI voice-phishing callers from their words, not their voice, with three detectors tested on an attacker AI they have never seen.
Challenges
Designing a fair test, so the detectors are judged on an attacker they were never tuned for, and collecting the data inside the fair's 12-month window.
Skills used
- Python
- PyTorch
- Transformers
- Machine learning
- Research design
Current status
In the planning stage. The research design is written and the pilot has not started.
Brief
Vishing is voice phishing, where a scammer phones you. AI can now fake the voice, so a detector that listens to the voice is easy to fool. My project catches these callers from what they say instead.
Plan
I compare three detectors. One is a simple keyword detector. One needs no training and scores how predictable the words are with GPT-2. One is a DistilBERT model that I fine-tune myself. I test all three against an attacker AI none of them has seen, then run them live on a Jetson Orin Nano.
This is for Scientific and Engineering Research at North Gwinnett High School. I am aiming at the ISEF Cybersecurity category and the NSA Research Directorate special award.
Process
I keep a dated logbook of every work session and a disclosure note for every use of an AI tool. I write the research plan, abstract and poster myself.