Untargeted Attack on Targeted-label for Multi-label Image Classification

Multi-label Image Classification (MLIC) is widely used in scene understanding, multi-object recognition, visual question answering and so on. Although providing promising performance, CNN-based models have been shown to be highly vulnerable to adversarial examples. Besides the targeted and untargeted attacks, which are introduced in Single-label Image Classification, there is a new type of attack on MLIC, i.e., the untargeted attack on targeted-label. In the untargeted attack on targeted-label for MLIC, one tries to interfere with the prediction of CNNs, causing untargeted attack on a targeted-label. To achieve this attack, we propose to reduce the spatial range of the adversarial attack from the whole image to where the object of the targeted-label is most likely to appear. We conducted the untargeted attack on targeted-label for MLIC in both PASCAL VOC 2007 and MS-COCO 2014 dataset. Experimental results indicate that the proposed method is effective to undermine CNN-based models for MLIC. The success rate of the untargeted attack on targeted-label using the proposed mask increases about 10%. In VOC2007, the success rate increase from 67.9% to 75.7%. In COCO2014, the success rate increases from 55.7% to 66.9%.