Datasets used to reproduce BadMIM

This dataset includes all necessary data to reproduct the proformance of BadMIM on CIFAR10.Specifically, for training auxiliary models, auxiliary datasets are provided for each target class in CIFAR10, which are constructed using Caltech256 and web-sourced target-class images. Notably, the \dog\ class already exists in Caltech256, so the auxiliary dataset for attacking the \dog\ class contains only Caltech256 images. For trigger augmentation (Adversarial Trigger Augmentation), ten triggers corresponding to each target class were adversarially augmented using their respective auxiliary models and datasets to produce the final augmented triggers.For backdoor injection (Reconstruction Hijacking), a shadow dataset (Imagenette2) and the augmented triggers are used to train a backdoored MIM encoder. This encoder is then fine-tuned on downstream tasks, such as CIFAR10.