Unveiling the Tricks: Automated Detection of Dark Patterns in Mobile Applications

Accepted to UIST2023

RESOURCE

• Paper: ACM
• Talk: Youtube
• Dataset: Zenodo
• Code: src
• FRCNN Pretrained Models: Zenodo

CODE

All code is tested under Ubuntu 16.04, Cuda 9.0, PyThon 3.9, torch 1.12.1, Nvidia 1080 Ti and also tested under MacOS 13.2.1, Apple M1 Pro

Usage

Usage: python3 UIGuard.py

UIGuard.py: code for detecting deceptive patterns

Modify L164 test_data_root to your data stored path
Modify L177 parameter _vis_ to decide whether draw the result or not
Update the path to the classification models in iconModel/get_iconLabel.py and statusModel/get_status.py
Update the path to the templates in template_matching/template_matching.py

rule_check.py: examine the dark patterns existence based on the extracted properties

L29-L32
flag_icon = True
flag_TM = True
flag_status = True
flag_grouping = True

Modify them to choose whether to use icon information (flag_icon) for examination.
Similar to other flags

evaluate.py

Evaluate the detected dps against the groudtruth dark patterns
Output some metrics(precision, recall, F1)

If you have any configuration problems with Faster RCNN, please refer to https://github.com/chenjshnn/Object-Detection-for-Graphical-User-Interface.