RL-PDNN: Reinforcement Learning for Privacy-Aware Distributed Neural Networks in IoT Systems

Due to their high computational and memory demand, deep learning applications are mainly restricted high-performance units, e.g., cloud edge servers. Particularly, in Internet of Things (IoT) systems, the data acquired by pervasive devices is sent computing servers for classification. However, this approach might not be always possible because limited bandwidth privacy issues. Furthermore, it presents uncertainty terms latency unstable remote connectivity. To support resource delay requirements such paradigm, joint real-time co-inference framework with IoT synergy was introduced. scheduling distributed, dynamic Deep Neural Network (DNN) inference requests among resource-constrained has been well explored literature. Additionally, distribution DNN drawn attention protection sensitive data. In context, various threats have presented, including white-box attacks, where malicious can accurately recover received inputs if model fully exposed participants. paper, we introduce a methodology aiming at distributing tasks onto system, while avoiding reveal We formulate an as optimization problem, establish trade-off between co-inference, data, resources devices. Next, due NP-hardness shape our reinforcement design adequate highly namely RL-PDNN. Our system proved its ability outperform existing static approaches achieve close results compared optimal solution.