MULDEF: Multi-model-based Defense Against Adversarial Examples for Neural Networks.

Despite being popularly used in many application domains such as image recognition and classification, neural network models have been found to be vulnerable to adversarial examples: given a model and an example correctly classified by the model, an adversarial example is a new example formed by applying small perturbation (imperceptible to human) on the given example so that the model misclassifies the new example. Adversarial examples can pose potential risks on safety or security in real-world applications. In recent years, given a vulnerable model, defense approaches, such as adversarial training and defensive distillation, improve the model to make it more robust against adversarial examples. However, based on the improved model, attackers can still generate adversarial examples to successfully attack the model. To address such limitation, we propose a new defense approach, named MULDEF, based on the design principle of diversity. Given a target model (as a seed model) and an attack approach to be defended against, MULDEF constructs additional models (from the seed model) together with the seed model to form a family of models, such that the models are complementary to each other to accomplish robustness diversity (i.e., one model's adversarial examples typically do not become other models' adversarial examples), while maintaining about the same accuracy for normal examples. At runtime, given an input example, MULDEF randomly selects a model from the family to be applied on the given example. The robustness diversity of the model family and the random selection of a model from the family together lower the success rate of attacks. Our evaluation results show that MULDEF substantially improves the target model's accuracy on adversarial examples by 35-50% and 2-10% in the white-box and black-box attack scenarios, respectively.