This repo contains code for final project paper of ECSE 4964. The proposed method combines and involves two main components: Extracting ground truth labels before the reconstruction process and optimizing gradient using a cost function based on cosine similarity.
The script contains a sample reconstruction for three methods: The traditional DLG, improved DLG (iDLG), and iDLG with cosine similarity (our proposed method).
All three methods were configured identically using the LeNet model, the CIFAR-100 dataset, a learning rate of 0.1, 300 iterations, and the LBFGS optimizer.
Set up your Python environment and install the required packages. Then run the script.
Yixuan Zeng, Yifan Wang
Department of Electrical, Computer, and Systems Engineering
Rensselaer Polytechnic Institute
Troy, NY 12180