Reproduce “Privacy-Aware Joint DNN Model Deployment and Partitioning Optimization for Collaborative Edge Inference Services” on OpenResearch

About this paper

Zhipeng Cheng, Xiaoyu Xia, Hong Wang, Minghui Liwang, Ning Chen, Xuwei Fan, Xianbin Wang

Researchers from Soochow University, RMIT University, Tongji University, and collaborators developed an optimization framework for collaborative deep neural network inference at the edge, jointly managing model deployment, user-server association, and model partitioning. The system minimizes average inference delay while adhering to long-term privacy constraints, maintaining robust performance across varied resource and demand scenarios.

Key insights

• Jointly optimizing DNN model deployment, user-server association, and partitioning is critical for effectively balancing inference delay and privacy preservation in dynamic edge environments.
• Lyapunov optimization provides an effective mechanism for managing long-term stochastic objectives and constraints without requiring prior knowledge of future system states, adapting well to real-world edge dynamics.
• An explicit trade-off exists between minimizing inference delay and maintaining privacy, which can be controlled by a tuning parameter (alpha), allowing the system to prioritize either objective while ensuring long-term constraint satisfaction.

Summary by alphaXiv.