Abstract: Geometric deep learning models have been applied in 3D shape retrieval task successfully, and their security evaluation is also drawing the attention of researchers. This paper proposed a method of multi-view universal perturbation attack (MvUPA) for 3D shape retrieval evaluation, so as to generate perturbed samples with a higher success rate of attack. Firstly, a multi-view depth panoramic map-based network was designed to train an efficient embedding representation for multi-view 3D shape retrieval. Secondly, a fusion loss function and its attack mechanism beneficial to multi-input UPA was proposed. The loss function combined triplet loss and label loss, thereby improving the perturbation generation for different categories of samples with similar topology and same category samples with different topology. The experiments validated the attack effectiveness and stability of MvUPA on multi-view retrieval models. MvUPA brought the decrease rate (DR) up to 94.52%, and the DR of the fused loss function was about 3.0%–5.5% higher than that of a single loss function.  

Key words: 3D shape retrieval, multi-view universal perturbation attack, universal perturbation attack, geometric deep learning, fusion loss