Detecting Deep Fakes with Mice
summary
States that humans hit a median 88% and current algorithms about 92% on the same dataset, while a spectral bicoherence+SVM approach reaches 95% on LJ Speech; prototypes use WaveNet, Tacotron2, DC TTS, WaveGAN discriminators and the ASVspoof set. Builds a mouse behavioral task plus widefield calcium imaging that achieves ~75% training accuracy and only ~10% drop on novel tokens, and shows that enzyme digestion of peri‑neuronal nets restores learning in mice that previously fail.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 2:18Todays Fakes
- 3:06The Dangers
- 4:05Informal Study
- 5:28Machine and Biology
- 6:02What is a Deep Fake
- 7:48Making Good Deep Fakes
- 8:38Wavenet
- 9:06History of TTS
- 11:26Gann Discriminator
- 14:58References
- 15:40What Kind of Deep Fake Detection
- 16:31Listening to Speech
- 17:19CoArticulation
- 18:35Deep Fakes
- 19:56Mice
- 20:23Mouse
- 22:26Why Mice
- 23:38synapses
- 24:12reset synapses
- 24:38local inhibitory neurons
- 25:41summary
related talks
47mFighting AI with AI
31mOff-the-shelf HuggingFace models for audio deepfake detection
27mTruth or Fiction The Art of Unveiling DeepFakes in Text, Audio, Image
28mFace spoofing detection: theory and practice
26mDeepfakes, audiofakes, and the future of media
32mHeart to Heart and Eye to Eye with Deep Fakes
37mRecent Solutions Using Deep Learning Models/ The How and Why
34mNeural networks: insane in the membrane