Edge-Efficient Deep Learning Models for Automatic Modulation Classification: A Performance Analysis

TL;DR

The paper tackles the challenge of deploying deep learning-based AMC on edge devices by systematically evaluating pruning, quantization, and knowledge distillation, along with two combined strategies (Distilled pruning and Distilled quantization). Using three CNN architectures (VTCNN2, ResNet, InceptionNet) and the RadioML2016.10A dataset, it demonstrates that aggressive sparsity (up to ~98%) and substantial storage reductions (up to ~133x) can be achieved with minimal or even improved accuracy, especially when KD guides the compression stage. KD can boost performance for smaller students, while the combined DP and DQ methods provide a practical path to compact, fast, edge-ready AMC models. Overall, the work offers actionable guidance for deploying high-performance, resource-efficient AMC systems on edge devices and motivates future hardware-aware optimizations.

Abstract

The recent advancement in deep learning (DL) for automatic modulation classification (AMC) of wireless signals has encouraged numerous possible applications on resource-constrained edge devices. However, developing optimized DL models suitable for edge applications of wireless communications is yet to be studied in depth. In this work, we perform a thorough investigation of optimized convolutional neural networks (CNNs) developed for AMC using the three most commonly used model optimization techniques: a) pruning, b) quantization, and c) knowledge distillation. Furthermore, we have proposed optimized models with the combinations of these techniques to fuse the complementary optimization benefits. The performances of all the proposed methods are evaluated in terms of sparsity, storage compression for network parameters, and the effect on classification accuracy with a reduction in parameters. The experimental results show that the proposed individual and combined optimization techniques are highly effective for developing models with significantly less complexity while maintaining or even improving classification performance compared to the benchmark CNNs.

Figures (4)

• Figure 1: Performance of network pruning method (NT) for different values of $\epsilon$.
• Figure 2: Performance of model quantization method (PQ) with and without retraining for different values of $P$.
• Figure 3: Performance of the KD method for all three combinations as given in Table \ref{['tab:kd_cases']} with fixed temperature $T =10$.
• Figure 4: Comparison of performance of the proposed combined optimization strategies with benchmark VTCNN2.