Robustness of Multiple Classifier Systems with different fusions to evasion attack

Pattern recognition system is frequently applied in many security applications, for example, intrusion detection, face recognition, spam mail filtering, etc. These applications may potentially face different kinds of attacks from an adversary. Some studies showed Multiple Classifier Systems (MCSs) is more robust than single classifiers under an adversarial attack. This paper investigates the robustness of multiple classifier systems using different fusions of continuous-valued outputs on an application with Boolean feature experimentally. Our results suggest that the mean method is the most robust.