Portfolio
Created as part of my Dissertation: A user-centred deepfake detection prototype that combines machine learning models with an intuitive interface to help users evaluate the authenticity of digital media.
Created as part of my Dissertation: A user-centred deepfake detection prototype that combines machine learning models with an intuitive interface to help users evaluate the authenticity of digital media.
As my dissertation is over, this piece of development is done for now. Maybe I'll reopen it down the line, but for now my focus is elsewhere.
On-device image analysis using ONNX Runtime Web and WebAssembly
Choice of MesoNet-inspired CNN, LightweightNet or a weighted ensemble combining both models
Risk scores, confidence indicators and colour-coded result summaries
Adjustable detection sensitivity
Basic, additional-context and full-detail explanations
Drag-and-drop uploads, file selection and batch processing
Limited video support through captured frames and poster images
In-page media detection controls and result overlays
Popup interface for quick controls and access to the full analysis dashboard
Dedicated dashboard, media analysis, history, analytics and settings views
Searchable analysis history with risk filtering and sorting
Re-analysis of previously analysed media where a stored preview or source is available
User feedback options to agree, disagree or express uncertainty about results
Analytics covering processing time, confidence distribution, risk breakdown and feedback
Local persistence of settings, analysis history and media previews
JSON export of analysis results and evaluation data
Platform: Chromium browser extension using Manifest V3
Frontend: HTML, JavaScript ES modules and SCSS
Build tooling: Vite with compiled styles and bundled runtime assets
Inference runtime: ONNX Runtime Web using CPU-based WebAssembly execution
Media preparation: Browser-based image decoding, resizing and normalisation, with frame extraction for supported video workflows
Storage: Chrome extension storage for settings and results; IndexedDB for media previews
Communication: Chrome runtime messaging between the popup, dashboard, content scripts and background service worker
Observability: PostHog analytics controlled by the anonymous-analytics preference, alongside configured Sentry error reporting
Testing: Vitest and JSDOM with mocked Chrome extension APIs
Framework: PyTorch for model development, training and ONNX export
Models: MesoNet-inspired and lightweight convolutional neural networks
Ensemble: Weighted combination of outputs from both models
Data preparation: Python preprocessing tools for images, video frames and FaceForensics-style datasets
Evaluation: Training curves, model-comparison scripts and exportable evaluation metrics
Interpretation: User-facing explanations describe model estimates and decision thresholds to support informed judgement
Layered extension architecture: Separate popup, dashboard, content-script and background-service responsibilities
Centralised inference: The background service worker manages model loading, inference requests, scoring and result persistence
Modular interface: Dashboard pages use dedicated components for uploads, history, analytics, settings and re-analysis
Local media processing: Core detection runs within the browser, without a remote inference service
Persistent local data: Settings and results remain available between sessions through extension storage and IndexedDB
Separate research pipeline: Python training and evaluation tools produce ONNX models for browser deployment
The project was developed as an MSc dissertation prototype. Video support is frame-based, and detection scores are model estimates rather than proof of media authenticity.
