Adversarial threat modeling
As described in the context of the arms race problem, it is usually a best practice to conduct threat analysis in a proactive way, such that security practitioners play the role of an adversary and perform a comprehensive analysis of the learning models. Therefore, we will show the threat modeling in the context of AML, as well as the taxonomy of various sorts of adversarial attacks on ML models, in this section.
Attacker model
In the context of adversarial learning, it’s important to consider an adversary as a smart and adaptable entity that interacts with the learning system. This adversary has the ability to make changes to the data that the system learns from. These changes can be made either during the system’s training phase (when it’s learning and developing its abilities) or during the testing phase (when its performance is being evaluated). Such alterations by the adversary can shift the way the original data is structured...
