Adversarial Attacks and Defenses in Capsule Networks: A Critical Review of Robustness Challenges and Mitigation Strategies

Capsule Networks (CapsNets) have gained significant attention in recent years due to their potential for improved representation learning and robustness. However, their vulnerability to adversarial attacks poses challenges for their deployment in safety-critical applications. This paper provides a critical review of the robustness challenges faced by CapsNets and explores various mitigation strategies proposed in the literature. The review includes an analysis of the adversarial attacks targeting CapsNets, such as manipulating primary capsule votes and direct targeting of CapsNets’ votes. The computational cost of applying existing attack methods designed for Convolutional Neural Networks (CNNs) to CapsNets is also examined. To enhance the robustness of CapsNets, the incorporation of detection-aware attacks and innovative defense mechanisms is discussed. The effectiveness and efficiency of these defense strategies are evaluated through extensive experiments. The findings reveal the superiority of certain defense mechanisms in mitigating adversarial attacks on CapsNets. However, it is acknowledged that further research is needed to explore more robust attacks and approvals and to compare the robustness of CapsNets with CNNs. This critical review aims to provide insights into the current state of adversarial attacks and defenses in Capsule Networks, facilitating future research and development in this field.