A minimalist AI-powered system to detect Deepfake and AI-generated content from images, audio, and video. Built for reliability, explainability, and real-world deployment.
Flask : PyTorch : CNN (LightCNN / ResNet / Xception) : OpenCV : Librosa : Redis : Celery : PostgreSQL : AWS S3 : Docker
Find Face · Detect and crop the face from the image
Preprocess · Resize, normalize, enhance artifacts
CNN + Attention · Deep CNN (Xception / ResNet) with attention to detect texture & edge anomalies
Score · Generate a “real vs fake” confidence score
Explain · Highlight suspicious regions (eyes, neck, ears)
Extract Frames · Sample multiple frames (20–30 frames)
Detect Faces · Run face detection on each frame
Image-style Check · Apply image model on each face → per-frame scores
LSTM / Transformer · Analyze temporal inconsistencies (blinking, lip sync, motion)
Video Score · Aggregate into final verdict + highlight suspicious segments
Load Audio · Normalize input voice signal
Spectrogram / MFCC · Convert audio into 2D representation (mel-spectrogram / MFCC)
CNN · Detect abnormal frequency patterns
LSTM · Analyze temporal rhythm, prosody, unnatural smoothness
Final Score · Classify as real or deepfake voice + highlight suspicious time segments
git clone <repo>
cd DivyaDrishti
python -m venv venv
source venv/bin/activate # windows: venv\Scripts\activate
pip install -r requirements.txt
python app.pyApp → http://localhost:5000
docker-compose up --build