A Defense Framework for Privacy Risks in Remote Machine Learning Service

<div>Comparing the training/test accuracy of original training data (no perturbation) with adversarial perturbation training data (OPTMARGIN, AdvGAN, and FGSM). (a) OPTIMARGIN perturbation with constraint condition. (b) AdvGAN perturbation with constraint condition. (c) FGSM perturbation with constraint condition.</div>

Security and Communication Networks

fig5

Figure 5

Figure 5: A Defense Framework for Privacy Risks in Remote Machine Learning Service 