Back to projects

Case Study

Audio-Video Deepfake Detection System

Multimodal AI system for identifying manipulated videos with spatial and temporal features.

Architecture

Video frames are preprocessed before passing through the detection pipeline.

Spatial and temporal features are fused for classification and confidence scoring.

A Flask API exposes the model for real-time inference and integration.

Screenshots

Screenshot placeholder
Inference
Multimodal prediction flow and confidence scoring.

Challenges

  • Combining temporal and spatial signals without creating a fragile pipeline.
  • Keeping inference fast enough for practical analysis.
  • Returning a score that is useful rather than just a label.

Benchmarks

92% classification accuracy
Real-time inference
Explainable scoring

Lessons Learned

  • Strong preprocessing often matters as much as model choice in media problems.
  • Explainability improves trust when the output is meant for decision support.

Tech Stack

PyTorchOpenCVFlaskVision TransformersCNNs

Key Results

92% classification accuracy
Real-time inference
Explainable scoring