Deepfake Live Detector
Real-time AI detection of synthetic images and audio from a live webcam or microphone.
Deepfake Live Detector
Real-time AI-powered deepfake detection for images and audio, with live webcam and microphone capture.
View CodeProject Overview
Synthetic media is cheap to generate and expensive to catch by eye, so this project treats detection as a measurement problem rather than a single-model guess. An EfficientNet-B4 backbone scores images and a CNN-LSTM over Mel-spectrograms scores audio, and a gated mixture-of-experts layer fuses the two into a single verdict. Around that sit a 10-stage forensic pipeline, a FastAPI service, a Streamlit dashboard for camera, microphone and upload, and concurrent batch scanning.
The Challenge
A single classifier gives a number with no explanation, which is useless when the answer is going to a person deciding whether to trust a clip. Image and audio forgeries also leave different traces, and a model good at one is often poor at the other.
Our Approach
Two specialised models, one fusion layer, and ten classical forensic stages that run first: file inspection, provenance, EXIF metadata, quality, and spectral analysis. When something is flagged the system reports the measured features behind the decision — FFT frequency distribution, edge sharpness, texture uniformity, spectral centroid and rolloff — rather than a bare score.
Technology Stack
Key Features
Detection that is live, dual-modal, and willing to show its working.
Live capture
Thread-safe OpenCV camera capture and a rolling-buffer microphone stream, so detection runs on content as it happens rather than on an upload.
Dual-modal detection
Separate image and audio detectors, fused by a gated mixture-of-experts layer that weighs each signal by how much it should be trusted.
10-stage forensic pipeline
File inspection, provenance, EXIF metadata, quality, and classical signal analysis run alongside the neural models to catch what a model alone would miss.
Explainable results
Every flag comes with the measurements behind it — spectral centroid, MFCC formants, zero crossings, edge sharpness — so a verdict can be checked instead of trusted.
Batch and API modes
A concurrent batch processor scans whole directories, and a FastAPI service exposes the same
pipeline over /detect and /health.
Portable weights
One-command ONNX export moves the trained models off PyTorch and onto lighter runtimes.