Application of complex systems in neural networks against Backdoor attacks

Abstract

Through the success of artificial neural networks (ANNs) in different domains and their increasing computational complexities, third parties and MLaaS (machine learning as a service) has been vastly used to do the training procedure. Hence the high possibility for malicious training recently caused intense researches centered on making these ANNs robust against various types of attacks such as backdoors. Backdoor attacks makes the ANN to behave normally on clean data but causes targeted misclassification in presence of the trigger. In this paper we provide the first investigation about the influence of applying complex systems such as random and scale-free networks instead of fully-connected structures on the robustness of feed forward neural networks (FFANNs) against backdoor attacks.