Pushing Deepfakes to the Limit-Fake video calls with AI
summary
Shows a working pipeline that moves from DeepFaceLab-style face swaps to whole-head, real-time replacement, trained on many identities (N-on-N) and reinforced with GAN loss to fix creepy artifacts. Uses TensorFlow 2 for a scalable trainer and MediaPipe Face Mesh (468 landmarks) for robust alignment. Demos run at ~2 FPS on a MacBook but hit real-time on a 3090-class rig; a live Zoom spoof convinces someone to transfer money. Implementation is proprietary, not open source.
AI-written summary · about
chapters (16)
- 0:00Deepfakes introduction
- 3:57Why explore deepfake tech
- 8:27Deepfakes 2018 to 2019
- 13:45How deepfakes work
- 20:00GAN training demo
- 27:03Real-time limits
- 33:43MediaPipe & TensorFlow
- 37:56Live deepfake calls
- 42:13Cybersecurity risks
- 44:13Spotting deepfakes
- 45:38Future of deepfakes
- 47:43Ethics & society
- 49:26Q&A: Detection reliability
- 51:56Q&A: Tattoos/scars
- 53:19Q&A: Discriminators
- 54:45AI slide decks preview
related talks
01h 04mPushing Deepfakes to The Limit - Fake Video Calls With AI
51mFaceswap: Developing a platform for practical Deepfakes
56mDeep Fakes 2.0 - How Neural Networks are Changing our World
51mPlaying Offense and Defense with Deepfakes
16mThe Threat of Deepfakes to Remote Identity Verification Systems
32mHeart to Heart and Eye to Eye with Deep Fakes
38mDIY Deep Fakes
01h 03mIt's All About AI