NEURAL ATTENTION DISTILLATION: ERASING BACKDOOR TRIGGERS FROM DEEP NEURAL NETWORKS

Keyphrases

• Distillation 100%
• Deep Neural Network 100%
• Backdoor Trigger 100%
• Neural Attention 100%
• Teacher Networks 75%
• Backdoor Attack 75%
• Training Data 50%
• Student Networks 50%
• Model Performance 25%
• Model Prediction 25%
• Performance Degradation 25%
• Fool 25%
• Layer Attention 25%
• Defense Framework 25%
• Training Time Attacks 25%

Computer Science

• Deep Neural Network 100%
• Backdoors 100%
• Training Data 33%
• Performance Model 16%
• Model Prediction 16%
• Performance Degradation 16%
• Small Proportion 16%
• Attention (Machine Learning) 16%