Odysseus (Clean and Trojan Models)

A major reason for the lack of a realistic Trojan detection method has been the unavailability of a large-scale benchmark dataset, consisting of clean and Trojan models. Here we introduce Odysseus the largest public dataset that contains over 3,000 trained clean and Trojan models based on Pytorch.

While creating Odysseus, we focused on several factors such as mapping type, model architectures, fooling rate and validation accuracy of each model, and also the type of trigger. These models are trained on CIFAR10, Fashion-MNIST, and MNIST datasets. For each dataset, clean and Trojan models are trained for 4 different architectures. Namely Resent18, VGG19, Densenet, and GoogleNet for CIFAR10 and Fashion-MNIST and 4 custom-designed architectures for MNIST. We also considered various sources to target label mapping for the Trojan models.