AI-Driven Modulation Classification for Intelligent Covert Communication in Next-Generation Wireless Networks

Abstract

In the age of digital independence and the Ubiquitous computing era, characterized by pervasive intelligence and extreme reliability, data privacy, security in transactions, and stronger communication channels are mandatory requirements for designing communication systems. The conventional cryptographic algorithms are serving their best; however, an unconventional and more intelligent way of securing communication is required. Thus, 6G and above networks are utilising covert communication for intelligent data transfer, making it undiscovered in the case of autonomous systems and the Industrial Internet of Things(IoT) environment. However, these covert systems are facing issues like malicious data manipulation and data leakage. In this paper, we present a hybrid dual-stream neural network and transformer architecture for RF modulation signal classification. In-phase and Quadrature(IQ) samples from the RadioML2016.10a dataset and spectrograms generated using Short-Time Fourier Transform(STFT) were fed as input to the model by combining 1D and 2D convolutions with a transformer architecture. An attention-driven feature extraction approach was incorporated using transformer attention maps and evaluated against the interpretability of the Explainable AI features, though its contribution is small but statistically significant. This feature interpretability provides more critical insights in identifying the features that make the signal detectable, which in turn helps to design a stealthy covert communication. The proposed model has achieved a classification accuracy of 90.49% by effectively differentiating different modulation signals for the SNR $\ge 0$ dB and 53.55% classification accuracy across the full SNR sweep of 20 to $\mathcal{C}$ 18dB, evaluated on a held-out partition without retraining. Also, it is evident from the three-seed significance test that the self-attention mechanism provides a small but statistically significant accuracy gain when compared with no-attention fusion, as observed in the ablation study. Although the model’s accuracy does not exceed the SOTA baseline, its main contribution lies in interpretability, which is evaluated using a comparison between attention-based feature attribution and LIME, validated by an independently trained Warden, the binary detector. The false-alarm probability, detection probability, and ROC-AUC are the metrics used for evaluating the Warden detector. Thus, the proposed model bridges the gap between adaptive covert communication and modulation signal analysis for more privacy-preserving communication in 6G networks. Additionally, this model is a candidate fit for a digital twin of an adversarial receiver, adaptive modulation and real-time Artificial Internet of Things(AIoT) deployment.