Fishy

Simple app for classifying phishing content.

It uses:

• ealvaradob/bert-finetuned-phishing
• elixir-nx/bumblebee

Usage

• mix deps.get

• iex -S mix

• In Elixir shell:

• Fishy.run("Click this link to win a prize!")

or

• paste content into content.txt and run File.read!("lib/content.txt") |> Fishy.run()

Benchmark

:timer.tc(Fishy, :run, ["Click this link to win a prize!"])
{2097277,
 %{
   predictions: [
     %{label: "phishing", score: 0.9999933242797852},
     %{label: "benign", score: 6.691957423754502e-6}
   ]
 }}

Value 2097277 is in microseconds.

Support for other languages

Model bert-finetuned-phishing is a fine-tuned version of bert-large-uncased with this dataset: phishing-dataset. It supports only English.

To support other languages, multilingual BERT bert-base-multilingual-cased would need to be fine-tuned with phishing and non-phishing examples across different languages.

Instructions for fine-tuning with Bumblebee: https://hexdocs.pm/bumblebee/fine_tuning.html