Energy-Based Learning for Preventing Backdoor Attack

The popularity of machine learning has motivated the idea of Energy-Based Learning (EBL), which used Energy-Based Models (EBMs) proposed by Prof. Yann to capture dependencies between variables. In addition, the application of several machine learning tools into the field of backdoor becomes widespread as well. However, the current backdoor researches didn't consider the novel EBL tools. This paper studies both EBL methods and backdoor attack of machine learning. We propose an algorithm to leverage energy-based learning for preventing backdoor attack. Several case analysis in this paper has demonstrated the promising of applying energy-based learning to improve the backdoor protection techniques.