PhishNet

ML/LLM — Google × NUS phishing detection with unsupervised learning and a real-time Chrome plugin path.

Research · Google × NUS · 2023–2024

Catch phishing before users do

Research collaboration with Google (Sai Teja Peddinti) and Dinil Mon Divakaran (NUS) on automated phishing identification — unsupervised / one-shot learning, LLM-assisted semantics, and a real-time Chrome plugin path.

99.6%Reported accuracy*
1-shotLearning path
ChromePlugin prototype
LLMs+ vision
System PhishNet / PhishDynamics — a multi-signal phishing detector combining logo/visual similarity, deep representation learning (transformers, autoencoders, classifiers), LLM-assisted URL/page semantics, and a Chrome extension for real-time verification with server-side inference.
Problem Blacklists and supervised classifiers lag zero-day lookalikes. Labeling is expensive; attackers mutate domains and clone brand logos. Prior systems (list feeds, logo-matchers such as Phishpedia-style detectors) still struggle with annotation cost and novel clones.
Our contribution Built domain + visual similarity pipelines with reduced manual labeling; integrated LLM semantics; prototyped in-browser delivery; evaluated against strong baselines (Phishpedia, OpenPhish, VirusTotal-style feeds) with continuous adaptation to cut training-set churn.

Method

Layer Methods
Visual Logo / brand similarity, image transformers
Representation Autoencoders, one-shot / unsupervised learning
Language LLM-assisted page / URL semantics
Delivery Chrome extension + server-side inference

Achievements

  • Reported ~99.6% accuracy in the study setting vs strong baselines
  • Bridged academic phishing research and a productizable browser path
  • Collaboration with Google and NUS researchers

*Accuracy from collaboration / project evaluation notes — not a public Google product claim.