TOORNet: Training-Free Online Object Removal Network for Adversarial Attacks on Intelligent Video Surveillance Systems

Gitaek Lee, Taehwa Lee, Byeongjin Kim, Wooju Kim

Abstract: With the huge advance of deep neural networks (DNN) in recent years, modern video surveillance systems have often been equipped with DNN-based vision methods such as object detection, thus promoting operational efficiency. However, many adversarial attack methods to malfunction the DNN-based vision systems have also been introduced concurrently. In this paper, we suggest that object removal methods, which remove particular objects in the scenes by masking the objects and filling in the masked regions, can work as a new way of adversarial attacks to intelligent video surveillance systems that employ DNN-based object detection models. For this purpose, we propose a novel object removal network, Training-free Online Object Removal Network (TOORNet). Unlike most existing object removal methods and video inpainting ones, which are closely related to them, our method works online (i.e., processing a frame based on only the past frames right upon a model receives it) and in real-time, thereby enabling us to perform object removals in live streaming videos. Experimental results show that our method successfully removes particular objects in live streaming videos, hindering the intelligent video surveillance systems from detecting the objects.

External IDs:dblp:conf/miwai/LeeLKK24