Logo image
Device Assignment and Model Splitting Optimization for Resilient and Secure Multi-hop Split Learning
 

Device Assignment and Model Splitting Optimization for Resilient and Secure Multi-hop Split Learning

Dongyu Wei, Defeng Zhou, Yuchen Liu Mingzhe Chen
IEEE transactions on wireless communications, Vol.25, pp.1-1
2026-07-21
: https://doi.org/10.1109/TWC.2026.3713859
 
Contrastive learning deep embedded clustering Labeling Learning (artificial intelligence) Modeling Multi-hop split learning reinforcement learning semi-supervised learning Servers Training Vectors Algorithms Optimization Probability
In this paper, a resilient and secure multi-hop split learning (SL) over wireless network is investigated. In our model, multiple edge devices trains different parts of a large model, collectively contributing to the training of the full model, while some devices may experience accidents during training. To enhance model training resilience against device accidents while preventing eavesdroppers from intercepting model information, multiple devices can be used to train the same sub-model. However, assigning several devices to train the same sub-model will increase transmission delay and eavesdropping risks. This problem is formulated as an optimization problem whose goal is to maximize the successful completion probability of SL training while considering device selection, model splitting, and transmit power control. To solve this problem, we first introduce a semi-supervised clustering method that integrates contrastive learning and deep embedded clustering (DEC) method to determine the sub-models assigned to each device according to wireless network features and network topologies. Different from current supervised learning methods that require many labels obtained by exhaustive search, the designed clustering method only needs a few labeled data samples thus reducing the training cost. We then introduce an optimization based reinforcement learning (RL) framework that combines a monotonic value function factorisation (QMIX) framework for device transmit power optimization and a depth-first-search (DFS) method to determine feasible sub-model splitting positions. Simulation results show that the proposed method improves the SL successful training probability by up to 9% and 56% compared to value decomposition network (VDN) and independent Q-learning (IQL).
1
Logo image

Logo image