AttackVLM
AttackVLM copied to clipboard
'Chat' object has no attribute 'get_img_list'
Thanks for the interest. Please adapt our code of crafting adversarial samples shown in the main page to the training / inference code of a specific model as the principal ideas are the similar: Firstly we apply transfer-based attack through the visual encoder (it is often a pretraind Clip encoder), then we conduct the black-box attack on the LLM part of that VLM.
Dears, Have u ever fixed this problem? I found there's no effect on Minigpt-4 by using attacked photo which is generated by unidiffusion. I wonder whether my work goes wrong.