An indigenous detection capability.
DeepShield is a final year project built at Mehran University of Engineering and Technology. It detects face-swap and face-reenactment forgeries in video, and reports how confident it is rather than forcing a binary answer.
Synthetic video is now convincing enough to fabricate a public statement, impersonate an official, or manufacture evidence. Existing detection tools are foreign and closed: they cannot be audited, cannot be trusted with sensitive material, and are not maintained against locally relevant threats.
A detector that can be inspected, retrained and run entirely offline is a different kind of asset. That is what this project set out to build.
Only 103,169 of 303 million weights are ever updated.
The backbone is CLIP ViT-L/14. Rather than retraining the whole network, only its LayerNorm parameters are tuned.
This matters more than it sounds. Full fine-tuning teaches a model the specific artefacts of its training set, and it then collapses on forgeries made by any other tool. Preserving the pretrained representation is what keeps cross-dataset accuracy at 0.942 where CNN baselines typically fall to around 0.65.
Supervised by Dr. Amirita Dewani, Assistant Professor, Department of Software Engineering, MUET Jamshoro.
- Asfand Ali22SW041Model architecture, training and evaluation
- Jai Kumar22SW044Inference pipeline, web application, deployment
- Zarawar Khan22SW120Dataset engineering, benchmarking, robustness testing