计算机科学
对抗制
人工智能
判别式
代表(政治)
集成学习
熵(时间箭头)
标记数据
自编码
编码器
特征学习
边距(机器学习)
模式识别(心理学)
监督学习
训练集
数据建模
特征提取
机器学习
抓住
人工神经网络
深度学习
推论
数据挖掘
强化学习
特征(语言学)
最大化
先验与后验
外部数据表示
作者
Yilin Cai,Sheng Wu,Yuting Lei,Mingyuan Shao,Dingzhao Li,Shaohua Hong,Jie Qi,Haixin Sun
标识
DOI:10.1109/tccn.2026.3656401
摘要
Deep learning has revolutionized automatic modulation classification (AMC), yet its reliance on large-scale, closed-set datasets remains a critical bottleneck, as these models generally perform poorly when training data is scarce or when faced with open-set scenarios. Few-shot learning (FSL) offers a promising solution by enabling models to grasp the concepts of new classes using only a limited number of labeled samples. In this paper, we propose a self-supervised FSL framework with gradient-adversarial data augmentation (GADA), tailored for modulation classification tasks. Our framework integrates a Transformer-LSTM hybrid encoder pre-trained through selfsupervised learning on unlabeled base classes to capture universal temporal dependencies, combined with knowledge distillation and adversarial augmentation that generates gradient-guided perturbed samples to enhance feature robustness. Furthermore, we introduce two complementary ensemble strategies to enhance predictive performance under few-shot conditions. First, we propose a Mixture of Experts module augmented with shared experts (SMoE) to capture common knowledge, coupled with an auxiliary Exclusive Expert Entropy Loss (E3Loss) that encourages taskspecific experts to learn diverse and specialized representations. Second, we design a soft-voting ensemble that aggregates predictions from multiple independently fine-tuned classifiers, leading to more robust final decisions. Experimental validation on the RML2018.10a and RML2016.10a datasets demonstrates stateof- the-art (SOTA) performance, achieving 96.07% accuracy and 94.01% in standard FSL scenarios. Notably, our approach consistently outperforms conventional FSL baselines across various signal conditions, demonstrating a significant advantage. These results validate our framework’s ability to simultaneously address two critical challenges: reducing dependence on labeled data while maintaining robust discriminative power across diverse signal environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI