Research · AI · M.S. thesis, in progress
CMDF-Net
A cross-modal deepfake detection framework that fuses audio and visual signals. An adaptive routing layer detects which modalities are present and runs only the branches it needs, which keeps computation practical.
The problem
Deepfakes can be faked in the video, the audio or both, and not every input has both. A good detector should use the signals that are present and not pay for the ones that are missing.
How it works
- 01Video or audio input
- 02Modality detection and routing
- 03Video and audio branches
- 04Fusion layer
- 05Real or fake
Highlights
- Video, audio and combined paths across seven datasets: FaceForensics++, Celeb-DF-v2, ASVspoof 2021, Urdu Deepfake Audio, FakeAVCeleb, AV-Deepfake1M and DFDC
- Built in three phases: independent video and audio branches, a fusion layer, then modality routing
- Selective computation is positioned as a core contribution
Built with
More like this
- S&P 500 BenchmarkStatistical, ML and deep learning compared
- Digital Twin Grow Room AssistantAgentic monitoring and yield forecasting
- Financial Analyst CrewMulti-agent research app on CrewAI
Need something like this?
I build AI agents, automation and forecasting for small businesses and teams. see how I work with clients.