notesum.ai

Published at November 25

On the Reconstruction of Training Data from Group Invariant Networks

Released Date: November 25, 2024

Authors: Ran Elbaz¹, Gilad Yehudai², Meirav Galun³, Haggai Maron⁴

Aff.: ¹Technion - Israel Institute of Technology, Electrical and Computer Engineering; ²Center for Data Science, New York University; ³Weizmann Institute of Science; ⁴NVIDIA Research

Arxiv: http://arxiv.org/abs/2411.16458v1

Dataset	Group size	Train set size	AM	KKT	KKT + SAME-GD	KKT + DIP
MNIST	2	50	$0.482\pm 0.000$	$0.464\pm 0.000$	$0.429\pm 0.004$	$\mathbf{0.285\pm 0.018}$
MNIST	2	100	$0.484\pm 0.000$	$0.467\pm 0.000$	$0.450\pm 0.004$	$\mathbf{0.271\pm 0.000}$
CIFAR-10	2	50	$0.446\pm 0.000$	$\mathbf{0.346\pm 0.001}$	$0.369\pm 0.000$	–
CIFAR-10	2	100	$0.463\pm 0.000$	$0.371\pm 0.001$	$\mathbf{0.370\pm 0.007}$	–
MNIST	8	50	$0.490\pm 0.002$	$0.471\pm 0.000$	$0.465\pm 0.000$	$\mathbf{0.314\pm 0.021}$
MNIST	8	100	$0.494\pm 0$	$0.471\pm 0.001$	$0.469\pm 0.000$	$\mathbf{0.323\pm 0.025}$