A Defense Framework for Privacy Risks in Remote Machine Learning Service

Table 2

Performance of our defense framework.

Without defense	Train acc. (%)	Test acc. (%)	Attack acc. (%)	With defense	Train acc. (%)	Test acc. (%)	Attack acc. (%)
CIFAR	10 96.1	62.24	93.71	Defense framework (AdvGAN)	78.3	69.7	51.4
MNIST	100	57.9	89	Defense framework (AdvGAN)	91.66	91.12	52.5