Research Material for Minor Project: Deepfake Detection using Machine Learning

Introduction This blog presents the research and conceptual framework behind our minor project on deepfake detection using machine learning and neural networks. The methods discussed—including image forensics, metadata analysis, and layered video detection—are part of our study and exploration. These techniques have not yet been implemented but represent potential approaches we aim to investigate further.

Image Forensics with ELA (Error Level Analysis) ELA (Error Level Analysis) is a compression-based method that exposes tampered regions in an image by identifying inconsistencies in JPEG compression.

Working Process:

Resave the image at 90% quality (introduces uniform compression).
Compare the resaved image with the original.
Detect anomalies in compression levels across the image.

Why ELA Works:

Edited regions compress differently than untouched ones.
These are visualized through highlighted error level differences.
Helps isolate areas of manipulation effectively.

Metadata & MLP-Based Classification In addition to ELA, we used image metadata (EXIF) and a Multilayer Perceptron (MLP) to classify images.

Key Steps:

Extract metadata (device info, software, timestamps).
Preprocess ELA and metadata into feature vectors.
Train MLP to output a probability score for tampering likelihood. This hybrid model increases accuracy by combining visual and non-visual signals.

Multi-Level Video Deepfake Detection

Our project also explores asyntactic detection in videos, structured in three levels:

Detection Level	Indicators	Classifier Type
Physical Level	Blinks, lighting, motion blur	Weak Classifier (SVM)
Semantic Level	Lip sync, frame transition inconsistency	Strong Classifier (CNN)
Physiological Level	Facial tension, micro-expressions	MLP with context score

Download Research Material

These techniques have not yet been implemented but represent potential approaches we aim to investigate further.

Download PDF