Reuse Zoo

Reuse Zoo consists of 250 models for model reuse detection. This dataset includes 4 pre-trained models, 126 surrogate models, and 120 reference models with four popular classification model architectures, including ResNet18, ResNet50, MobileNet V2, and GoogLeNet. Four models collected from TorchVision are pretrained on ImageNet. For each pre-trained model, we use transfer learning method to obtain three surrogate models trained on CIFAR10, CIFAR100, and SVHN. These 12 models constitute the transfer learning surrogate set whose source victim models are the pretrained models. We construct eight surrogate models for each transfer learning reuse model by applying pruning, quantization, fine-tuning, and retraining reuse methods. We also include knowledge distillation and model extraction surrogate models for ResNet50 models. The 120 reference models are obtained by indepenently training ten models from scratch using different weight initialization for each model architecture and each dataset.