Vertically Federated Graph Neural Network for Privacy-Preserving Node Classification

Our paper received four reviews, including a ‘Very Strong Accept’ (9), ‘Strong Accept’ (8), ‘Accept’ (7), and ‘Borderline reject’ (4). The average rating of our paper is Accept (7). Unfortunately, we got reject at the final decision. The main concerns of the fourth reviewer are (1) the labels are not protected by differential privacy (DP), and (2) the given privacy loss is only for per iteration rather than the whole algorithm.

We have solved all the concerns raised by the reviewers, and the main changes and improvements are listed as below.

• •

  We modified the local node publishing mechanism and proposed a general information publishing mechanism using DP in Algorithm 1. We applied the information publishing mechanism in all the local information that the data holders send to server, including the local node embedding that calculated from private feature and edge, and the gradient update calculated from private label.

• •

  Besides the privacy loss of per iteration in Theorem 2, we further analyzed the privacy of the whole algorithm and presented it in Theorem 3. The proposed theorems prove that our proposed information publishing mechanism preserve $(\epsilon,\delta)$-DP.

• •

  We updated DP related experiments and reported the results in Table 5.