Towards Stable Backdoor Purification through Feature Shift Tuning

Keyphrases

• Backdoor 100%
• Poisoning 60%
• Backdoor Attack 40%
• Specific Model 20%
• Attacker 20%
• Small Sets 20%
• Comprehensive Evaluation 20%
• Training Samples 20%
• Entanglement 20%
• Behavioral Model 20%
• Real-world Application 20%
• Stable Performance 20%
• Training Process 20%
• Attack Scenario 20%
• Deep Neural Network 20%
• Defense Method 20%
• Tuning Methods 20%
• Vanilla 20%
• Parameter Adjustment 20%
• Model Driven Architecture 20%
• Lines of Defense 20%
• Classifier Weights 20%
• Tuning Cost 20%

Computer Science

• Backdoors 100%
• Attackers 14%
• Behavior Model 14%
• World Application 14%
• Training Process 14%
• Attack Scenario 14%
• Deep Neural Network 14%
• Training Sample 14%
• Model Architecture 14%