Backdoor Attacks on Machine Learning with Covert False Data Injection-Part B: Tests

Based on the content of Part A [11], in this paper, we conduct the FDI attack test based on the modeling. The test mainly focuses on backdoor attacks under two modes of partial FDI and full FDI, respectively, and we introduce compound evaluation metrics to assess the effectiveness of the attacks and test the attacks against the three concealed FDI modes of adversarial backdoors. The experimental test results show that the proposed Clandestine Fraudulence is the most destructive to performance, while Clean Label Attack is the weakest, and Induced Model Attack has the best accuracy destruction concealment. This work can provide a basis for the security design of machine learning system attack and defense and training algorithms.