Unveiling the secrets of a machine learning model
summary
Describes a GAN-based model inversion method that starts from a name vector and optimizes a latent code by backpropagating through a frozen generator and critic, requiring white-box access to the target network; reports measured success rates of 28–46% (up to 82% when using masks) on public datasets and notes the original implementation in PyTorch with efforts to port to TensorFlow. Evaluates differential privacy in training, showing DP-SGD with a configurable noise multiplier and ε/δ trade-offs reduces membership leakage but lowers accuracy and increases compute cost.
AI-written summary · about
chapters (5)
- 0:00Start
- 14:38Privacy analysis
- 23:011 Retrieve target output vector
- 25:033 Optimize portrait generation
- 41:57Conclusion
related talks
40mTraining generative models from locally privatized data via entropic optimal transport
01h 00mTutorial on Deep Learning IV
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
52mDeep Generative Models And Unsupervised Methods For Inverse Problems
30mPrivate Generative Models: Balancing Privacy and Utility in AI
45mSchemes, Privacy, and Security
01h 28mPPML: Machine Learning on Data you cannot see
40mOverview of Generative Adversarial Networks (GANs)