README.md

metadata

title: Fintech Dark Patterns Detector
emoji: 🛡️
colorFrom: blue
colorTo: gray
sdk: docker
app_port: 8000
pinned: false

Dark Pattern NLP Detector

This project detects potentially manipulative language in fintech interfaces. It supports screenshot auditing through OCR, direct text classification, model evaluation, dataset exploration, and a Chrome Manifest V3 extension.

NLP pipeline

1. Screenshot text is extracted with Tesseract OCR. Low-confidence OCR words are discarded before classification.
2. Text is represented using word TF-IDF n-grams (1, 3) and character char_wb n-grams (3, 5).
3. A class-balanced Logistic Regression model predicts one of seven dark pattern classes or Not Dark Pattern.
4. Sigmoid calibration converts raw model scores into better-behaved confidence estimates.
5. Word-level linear contributions provide an explanation for each prediction.

Evaluation uses a grouped held-out fold based on page_id. This prevents text from the same source page appearing in both training and testing.

Run locally

npm install
npm run build
cd ai-backend
python server.py

Open http://127.0.0.1:8000.

Tesseract OCR must be installed. On Windows, the backend automatically checks C:\Program Files\Tesseract-OCR\tesseract.exe.

Test

npm run test:backend
npm run check:extension
npm run build

The backend tests cover grouped evaluation, calibrated text classification, batch IDs, remote-image blocking, and the screenshot OCR response.

Chrome extension

The ready-to-load extension is in extension. Start the local API, open chrome://extensions, enable Developer mode, choose Load unpacked, and select the extension directory itself. The selected directory must contain manifest.json at its top level; do not select the ZIP or the space directory.

Generate both a ZIP and a Chrome-ready unpacked artifact with:

npm run package:extension

For the least ambiguous installation path, select artifacts\dark-pattern-nlp-auditor-unpacked in Chrome after running that command.

The extension reads visible DOM text only after the user clicks it. It sends one batch request to the configured local API, highlights flagged elements, and shows results in Chrome's side panel.

Known limitations

• Forced Action, Sneaking, and Obstruction have fewer than 100 labeled examples, so their class-level metrics remain unreliable.
• OCR cannot perfectly recover text from low-resolution or highly stylized UI.
• Closed shadow roots, cross-origin frames, canvas, and image-only web content are not available to the extension's DOM scanner.
• Compliance descriptions are screening guidance, not legal conclusions.