A Lightweight Cross-Feature Slicing Model for Real-Time Resource Allocation in 6G Edge Networks

Abstract

The 6G wireless networks emerge, the need for realtime, efficient resource allocation in Internet-of-Vehicles (IoV) systems becomes critical. Existing machine learning approaches achieve high accuracy, but produce models of several hundred kilobytes, making them unsuitable for deployment on IoT-class Multi-access Edge Computing (MEC) nodes. This paper proposes the Lightweight Cross-Feature Slicing Model (LCSM), a family of three unique neural architectures - LCSM-Base, LCSM-Nano, and LCSM-XAI that combine feature space attention, residual skip connections, and bottleneck compression to achieve high accuracy at minimal parameter counts. Evaluated against eight baseline models on a 6G IoV dataset comprising 5,000 samples and 8 engineered features. LCSM-Nano achieves 96.93% accuracy at just 7.57 KB (1,937 parameters), representing a 13× smaller deployed size than the CNN-BiLSTM baseline (94.93%, 98 KB) while exceeding its accuracy by 2.0 percentage points. Post-training INT8 quantization reduces LCSM-Nano to approximately 7.41 KB, and on-device benchmarking on a Raspberry Pi 4 confirms a mean inference latency of 0.070 ms and a throughput of 14,336 decisions sufficient for real-time 6G IoV slice management on severely constrained hardware. Ablation studies quantify each architectural component's contribution, and the LCSM-XAI variant provides inherent, per-sample attentionbased feature importance, directly extending the Explainable AI programme of prior work.