Moonash16/Final-year-project-CSC400

This a final year project titled "Developing a lightweight deepfake detection system for social media in Eswatini" # PreMune Secure **Developing a Lightweight Deepfake Detection Framework for Social Media in Eswatini** Final Year Project: BSc Information Technology, Department of Computer Science, University of Eswatini (Kwaluseni Campus) **Authors:** Precious T. Ndwandwe (202200294) & Munashe Matsanura (202203710) **Supervisor:** Mr. E. L. Dube ## Overview `PreMune Secure` is a lightweight, fully offline Android application designed to detect manipulated media (deepfakes) directly within the social media platforms where such content is most commonly encountered in Facebook, TikTok, and WhatsApp. It was built to address the lack of accessible, low-resource, on-device deepfake detection tools for regions like Eswatini, where computationally intensive detection solutions are impractical. The system combines a custom-trained, quantized deepfake detection model with a multi-metric forensic analysis suite (Error Level Analysis, noise profiling, and structured pattern detection), fused through weighted averaging for stronger detection accuracy. ## Key Results - **74.3%** validation accuracy (custom fine-tuned MobileNetV2 classifier) vs. **46.0%** for a generic pre-trained MobileNetV2 baseline representing a `28.3` percentage point improvement `INT8 post-training quantization` reduced the model's size by `89.2%`, shrinking it from `27.3 MB down to 2.96 MB`. The model was trained on a balanced dataset of 6,000 real and synthetic face images, and the final application runs on Android 5.0 and above through a persistent floating overlay service compatible with Facebook, TikTok, and WhatsApp. ## Repository Structure This repository contains the full working project, including the Android app source, model training scripts, trained models, and the final APK build, organised as a flat working directory rather than a separated package structure: The Android app source code includes files such as `mainactivity.java`, `overlayservice.java`, `deepfakedetector.java`, `forensicanalyzer.ja …