QSplitFL: Capability Aware Deep Q-Learning for Optimal Split Point Selection in Split Federated Learning

QSplitFL introduces a Deep Q-Network framework that optimizes split point selection in federated learning by considering device heterogeneity, using lightweight hardware metrics instead of model weights. The approach demonstrates improved convergence and accuracy across multiple datasets and neural network architectures while adapting to varying client capabilities.

QSplitFL addresses a fundamental inefficiency in Split Federated Learning (SFL) systems where model partitioning doesn't account for device diversity. Traditional approaches use fixed split points that create bottlenecks on resource-constrained devices, degrading training speed and stability. This research leverages reinforcement learning to dynamically determine optimal splitting strategies based on real-time hardware telemetry—CPU utilization, memory, battery, and network latency—rather than abstract model representations.

The broader context reveals a critical gap in distributed machine learning infrastructure. As edge computing adoption accelerates, federated learning becomes increasingly valuable for privacy-preserving training on heterogeneous devices. However, existing SFL implementations fail to optimize for real-world hardware constraints. QSplitFL's committee-based DQN architecture with majority voting introduces a novel safeguard against reward hacking, a known vulnerability in RL-based optimization that could otherwise distort training dynamics.

For practitioners and organizations deploying federated learning systems, this framework offers tangible benefits: faster convergence, reduced server load, and improved resource utilization across mixed-capability networks. The lightweight state representation makes the approach computationally efficient compared to weight-based alternatives, reducing communication overhead. Testing across MNIST, CIFAR-10, CIFAR-100, and multiple architectures (CNN, ResNet50, MobileNetV4, ConvNeXt) demonstrates robust generalization.

Looking ahead, the public code release enables broader adoption and validates whether capability-aware splitting becomes standard practice in federated systems. Future developments should explore how dynamic split selection performs under extreme heterogeneity and whether the approach scales to production-scale distributed training pipelines across thousands of edge devices.

• →QSplitFL uses hardware metrics rather than model weights to determine optimal split points in federated learning, improving efficiency across heterogeneous devices.
• →A committee-based DQN with majority voting mitigates reward hacking vulnerabilities inherent in reinforcement learning optimization.
• →The framework achieves faster convergence and higher accuracy while reducing server load and communication overhead compared to fixed split-point approaches.
• →Testing across four datasets and multiple neural network architectures demonstrates robust generalization to diverse training scenarios.
• →Public code release enables adoption and validation of capability-aware splitting as a standard practice in distributed federated learning.