Skip to content

Free · Open source · A CSIT375 student project

How Fishy is your link?

Paste a link or a message to find out if it’s phishing. Fishy scores it with two machine learning models, tells you why, and blocks it in your browser if the risk is high.

Try

Fishy never opens the links it checks.

Example result
DangerousLooks like phishingExample

Blocked automatically

http://paypa1-secure-update.com/signin

Model A says89.5%
Looks at 30 features of the address, like its length and symbols
Model B says95.8%
Looks at short letter patterns that often appear in phishing links

Why Fishy flagged it

  • Pretends to be PayPal. “paypa1” uses the digit 1 instead of the letter l.
  • Asks you to sign in. The address contains “signin”.
  • Hyphenated name. “paypa1-secure-update” uses hyphens, a common way to sound official.
  • No HTTPS. Anything you type would not be encrypted.
This is a real result for a made-up phishing link. Check a link above to see your own.

How does Fishy decide?

Two models score the link separately. Fishy averages the two scores, checks a short trusted list, and turns the result into one of three answers.

  1. 1. You paste a link or a message.For a message, Fishy finds every link inside it.
  2. 2. Two models score it.Model A (Random Forest) checks 30 features of the address. Model B (TF‑IDF with Logistic Regression) checks letter patterns.
  3. 3. The scores are averaged.The average becomes a risk score out of 100.
  4. 4. Then Fishy checks a short trusted list.If the site is a well known one like google.com, Fishy shows it as safe whatever the average, and still shows you both scores.
  5. 5. Fishy acts on the result.70 or more is blocked and you are alerted. 40 to 69 gets a warning.
Examplepaypa1-secure-update.com/signin
Model A89.5%
Model B95.8%
(89.5 + 95.8) ÷ 2 =93Dangerous
  • Safe0 to 39
  • Warning40 to 69
  • Blocked70 and over

How accurate is Fishy?

Each model is strong on the kind of data it learned from and weak on the other. Averaging them is never the top score, but its worst result is far better than either model’s worst.

Results from

In short: On links unlike their training data, Model A alone scored as low as 0.251 out of 1 and Model B as low as 0.632. Fishy’s average never dropped below 0.714.

Web-page dataset

1,847 links, F1 score

Model A
0.868
Model B
0.632
Both averaged (Fishy)
0.836

URL dataset

103,904 links, F1 score

Model A
0.251
Model B
0.765
Both averaged (Fishy)
0.714

Worst of the two

Model A
0.251
Model B
0.632
Both averaged (Fishy)
0.714
What do these numbers mean?

F1 combines how many phishing links were caught with how many safe links were wrongly flagged. 1.0 is perfect. Threshold 0.5. Any link that either model saw during training was removed first. On their own test sets the scores are higher: Model A 0.888, Model B 0.891.

See the code and raw results on GitHub
65.7%

Why did adding a “/” change the verdict?

In the first version of Model B, adding a slash to the end of a bare domain flipped the verdict 65.7% of the time, across 14,528 test domains. In the training data, bare domains were mostly phishing. We removed trailing slashes before training. Now it flips 0% of the time.

What happens when Fishy finds a dangerous link?

  1. 1. You get an alert right away.

    A notice on the page, plus a desktop notification if you turn them on.

  2. 2. It goes on your blocklist.

    If you check it again later, Fishy warns you without asking the models.

  3. 3. Opening it needs a second click.

    A warning explains the risk before you can continue.

  4. 4. You can take the list with you.

    Export it as a hosts file to block the sites on your whole computer.

fishy-blocklist.hostsExample export
# Exported from Fishy
# 2026-11-09
# Copy these lines into your hosts file.

0.0.0.0 paypa1-secure-update.com
0.0.0.0 free-gift-card.xyz
# 192.168.4.2 (an IP address, can't be blocked here)
Warning before openingThis link is blocked.Risk 93 out of 100. “paypa1” uses the digit 1 instead of the letter l.

What has Fishy blocked on this device?

Your blocklist, history and alerts. They are saved only in this browser and never sent anywhere.

Nothing blocked yet

Dangerous links you check land here automatically, so Fishy can warn you straight away next time. Exports turn on once something is here.

Checks paypa1-secure-update.com, a made-up phishing link.

What did we find that we didn’t expect?

Three results from testing our own models. Each one changed how Fishy works.

0.961 → 0.955

Two columns were giving the answer away.

“google_index” and “page_rank” come from outside services, not from the page itself. The dataset’s authors also found outside-service features the most predictive. We removed them so the test is fair, and Model A’s F1 barely moved.

Source: Hannousse and Yahiouche, 2021
1,284

training links were marked as IP addresses but had none.

For example “support-appleld.com.secureupdate…” is a normal domain name. We checked every feature against our own calculation and left out the three that did not match.

Source: our feature check, on GitHub
0.251

Model A’s F1 on the other dataset.

A model can score 0.888 on its own test set and still fail on links collected a different way. This is why Fishy uses two models and shows you both scores.

See the accuracy table

Questions people ask

What do “Random Forest” and “TF‑IDF” mean?

A Random Forest is a few hundred small decision trees that each vote, for example “is the address longer than 75 characters?”. TF-IDF counts short groups of letters, like “pay” or “.exe”, and gives more weight to the ones that are rare. Logistic Regression then turns those counts into a probability.

Does Fishy open the links it checks?

No. Fishy only reads the text of the address. It never visits the site, so checking a link can’t infect your device or tell the site that you looked.

Why did Model B give google.com 68.8%?

Model B learned from a dataset where most bare homepages, a name with nothing after it, were phishing pages. So it is wary of any short address. After scoring, Fishy checks a short list of well-known sites and marks them safe. For google.com this changes nothing: the average is 34, which is already Safe. Both scores are still shown.

What data were the models trained on?

Model A learned from 9,143 labelled web pages (the Web page phishing detection dataset). Model B learned from 512,890 labelled links (the Malicious URLs dataset). Both were cleaned and split so that no link appears in both training and testing.

Where are my checks stored?

Your blocklist, history and alerts stay in this browser and never leave it. The links you check are sent to Fishy’s server only to be scored. They are not saved or logged there.

Can I use Fishy instead of antivirus?

No. Fishy is a student project. On links unlike its training data it still misses some phishing and flags some safe sites, as the accuracy table shows. Use it as a second opinion, alongside your browser’s built-in protection.

Who built Fishy?

What
A phishing-link checker built for CSIT375 AI for Cybersecurity at the University of Wollongong.
When
November 2026
Code
Open source on GitHub
Results from

Sources

  1. Hannousse, A. and Yahiouche, S. (2021). Towards benchmark datasets for machine learning based website phishing detection. Engineering Applications of Artificial Intelligence, 104, 104347. arXiv:2010.12847
  2. Web page phishing detection dataset (shashwatwork). Kaggle
  3. Malicious URLs dataset (sid321axn). Kaggle
  4. scikit-learn documentation. scikit-learn.org

Got a link you’re not sure about?