A small scale multi-class phishing web page screenshots archive

Hello, this is the home page of Phish-IRIS dataset.

Phish-IRIS dataset is aimed for researchers to supply a ground truth dataset to evaluate their vision based multi-class anti-phishing studies. For this purpose, we supply a corpus involving unique screenshots of 15 (14+1) classes. Here, 1 class represents the "unknown" or "legitimate" samples while the rest of the 14 classes correspond to different highly phished brands. It is important to mention that, Phish-IRIS dataset aims to provide a benchmark dataset for only computer vision based anti-phishing studies.

Our dataset involves 1313 training and 1539 testing samples. The directory structure of the dataset has been formed in order to be used without an extra effort. In this regard, you can employ it in several Deep Learning frameworks such as Caffe©, Pytorch©, Tensorflow© and Keras©.

Since the nature of the anti-phishing is based on discriminating legitimate web pages from the phished targets, we also provided a fairly larger set for "legitimate" samples.

In this web page, we will publish the state of the art results obtained from Phish-IRIS dataset on a regular basis. Moreover, we are planing to list the scientific papers citing this dataset. Please keep an eye on "How to Use" section for more detailed information and the motivation behind the need and generation of this dataset.

This dataset has been generated in HUMIR LAB at Hacettepe University Department of Computer Engineering and it is intended to be used for only academic purposes. Please read the paper provided at the menu before proceeding.


State of the Art (SOTA)

In terms of TPR, FPR and F1 - Current SOTA: Dalgic, Bozkir and Aydos

True Positive Rate 90.6%
False Positive Rate 8.5%
F1 - Measure 90.5%

Details & How to use

1. The Phish-IRIS data set has been collected in order to involve both phishing and legitimate samples. Legitimate samples are collected under "other" directory while phishing samples are located inside of respective brand names. We have also seperated train and test folders at first branching and made it ready to be used along with contemporary deep learning frameworks

2. The image sizes ara varying. Therefore you may need to make resize operation before proceeding

3. In order to create a challanging data set, unlike other studies, we considered the data set to be as diverse as possible. So, you may notice that we have included many unique samples during data collection

4. Please acknowledge us if you really interest to our dataset, thus we can publish your scores in this web page

5. Cite our paper if you use our data set either for benchmarking or method development.