A collection of 840 pretrained VQA models which may be regular “clean” models or malicious “backdoored” models which have been trained to include a secret backdoor trigger and behavior. This collection includes models with traditional single-key backdoors as well as Dual-Key Multimodal Backdoors.
For more information, see our work “Dual-Key Multimodal Backdoors for Visual Question Answering" (https://arxiv.org/abs/2112.07668).
This dataset is inspired by and modeled after those created by TrojAI (https://arxiv.org/abs/2003.07233). It is intended to enable the development of defensive algorithms to detect and/or purify backdoored VQA models.
Paper | Code | Results | Date | Stars |
---|