bboylyg / keras-resources Goto Github PK

I am a research fellow at the School of Computing and Information Systems at Singapore Management University supervised by Prof. Jun Sun. I also work closely with the Prof. Xingjun Ma at Fudan university. I have completed my Ph.D. degree at Xidian University supervised by Prof. Xixiang Lyu. Research publications in Google Scholar.

• Understanding the effectiveness of backdoor attacks
• Robust training against backdoor attacks
• Design and implement a general defense framework for backdoor attacks

• Yige Li, Xingjun Ma, et al., “Multi-Trigger Backdoor Attacks: More Triggers, More Threats”, submitting, 2024.
• Yige Li, Xixiang Lyu, et al., “Reconstructive Neuron Pruning for Backdoor Defense”, ICML 2023.
• Yige Li, Xixiang Lyu, et al., “Anti-Backdoor Learning: Training Clean Models on Poisoned Data”, NeurIPS 2021.
• Yige Li, Xixiang Lyu, et al., “Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks”, ICLR 2021.

• Neural Attention Distillation (NAD)

  • A simple and universal method against 6 state-of-the-art backdoor attacks via knowledge distillation
  • Only a small amount of clean data is required (5%)
  • Only a few epochs of fine-tuning (2-10 epochs) are required

• Anti-Backdoor Learning (ABL)

  • Simple, effective, and universal, can defend against 10 state-of-the-art backdoor attacks
  • 1% isolation data is required
  • A novel stratrgy benefit companies, research institutes, or government agencies to train backdoor-free machine learning models

• [email protected]
• [email protected]