Adversarial Backdoor Attacks on Machine Learning with Covert False Data Injection-Part A: Modeling

In this paper, we focus on the adversarial machine learning in financial services. In recent years, an attack pattern of false data injection against AI computing has gradually emerged in the industry. The attack is low cost and well concealed, and the data injection will directly distort the rules of the rating system, so it is not easy to be detected. To achieve better learning results, machine learning often uses massive amounts of data for training. However, while improving the generalizability of the model, massive data also brings the labor and time cost of data review, which is not easily checked frequently. This also leads to false data injection (FDI) against machine learning which is often not easily detectable. In this paper, seven FDI-based adversarial backdoor attacks against machine learning systems have been proposed with analytic modeling. Combining with actual rating application scenarios, we discuss the patterns and stealth characteristics of each attack, focusing on three more stealthy backdoor attack patterns under FDI, including Clandestine Fraudulence, Clean Label Attack, and Induced Model Attack, and perform targeted modeling analysis. Based on the quantitative analysis, we parameterize the actual attacks and describe each backdoor attack 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. A case study with tests will be given in Part B.