Reconstructing Training Data from Model Gradient, Provably

Keyphrases

• Training Data 100%
• Training Samples 50%
• Parameter Values 25%
• Activation Function 25%
• Deep Neural Network 25%
• Tensor Decomposition 25%
• Computationally Efficient Algorithms 25%
• Shallow Neural Network 25%
• Federated Learning 25%

Computer Science

• Training Data 100%
• Training Sample 50%
• Efficient Algorithm 25%
• Mild Condition 25%
• Parameter Value 25%
• Information Leak 25%
• Federated Learning 25%
• Activation Function 25%
• Deep Neural Network 25%

Mathematics

• Training Data 100%
• Training Sample 50%
• Wide Range 25%
• Identifiability 25%
• Tensor Decomposition 25%
• Deep Neural Network 25%