Distributed Private Aggregation in Graph Neural Networks
summary
Claims a practical distributed GNN aggregation that attains node-level differential privacy by combining secure multi-party computation with three MPC-friendly optimizations (matrix addition, sparse key–value sharing, dummy keys) plus training changes that consume privacy budget once and reconstruct secrets on users. Implements DPA-GNN and evaluates on six real-world datasets, reports consistent wins versus baselines under tight budgets, reduced communication/offline costs, and explicit privacy accounting (ε_E, ε_H, ε_Y) in a semi-honest, anonymous-user model.
AI-written summary · about
chapters (12)
- 0:00Distributed private aggregation in GNNs
- 1:55Privacy challenges in GNNs
- 3:40Problem setting and threat model
- 4:44Core idea of DPA method
- 6:07Optimizing aggregation efficiency
- 7:34Reducing privacy budget oversplitting
- 8:58DPA function and subfunctions
- 9:45DPAGNN workflow
- 10:48Privacy guarantees and efficiency
- 11:54Experimental setup
- 13:20Conclusion and future work
- 14:03Q&A
related talks
33mPrivacy of Decentralized Machine Learning
30mDistributed Aggregation Protocol: A Standard for Privacy-Preserving Aggregation in MPC
35mDistributed Models for Private Analysis of Graph Data
34mEfficient Robust Private Aggregation for DP
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
31mScaling privacy embedded distributed ML systems
31mIs your model private?
21m"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving