CLIENT-ADAPTIVE NEURAL NETWORKS: A SERVER-DRIVEN ARCHITECTURE FOR DYNAMIC MODEL PARTITIONING IN EDGE COMPUTING

The deployment of deep neural networks (DNNs) on resource-constrained edge devices is a significant challenge in contemporary computing. A core problem is that traditional model partitioning, which splits a DNN at a fixed point between a client and a server, fails to account for the dynamic nature of edge environments, including fluctuating network bandwidth, heterogeneous client hardware, and varying computational loads. This paper proposes Client-Adaptive Neural Networks (CANN), a novel server-driven architecture for dynamic model partitioning. In CANN, a lightweight profiler on the client device continuously monitors resource metrics (CPU, memory, battery) and network latency. These metrics are reported to a central Orchestration Server, which leverages a reinforcement learning (RL) agent to dynamically select the optimal partition point from a set of pre-defined "bottleneck" layers within a DNN. This decision minimizes end-to-end inference latency while respecting client resource constraints. We evaluate our framework on a variety of DNN models (MobileNetV2, ResNet-50) and edge device simulators under realistic network conditions. Our results demonstrate that CANN reduces average inference latency by 41% compared to static client-side, server-side, and fixed-split approaches, and improves energy efficiency by 33% on client devices. We conclude that the CANN architecture provides a robust, adaptive solution that effectively overcomes the limitations of static partitioning, paving the way for more responsive and sustainable edge AI systems.