International Research Journal of Engineering and Technology (IRJET) Volume: 13 Issue: 08 | Aug 2026
www.irjet.net
e-ISSN: 2395-0056 p-ISSN: 2395-0072
Human-Centered AI Governance Framework for Ethical and Explainable Hybrid SDR Signal Intelligence Systems 1Diwanji Chaitanya Anant, 2Gaurav Sharma 1Faculty of Communication Engineering Military College of Telecommunication Engineering 2Senior Member, IEEE, Faculty of Communication Engineering Military College of Telecommunication Engineering
-----------------------------------------------------------------------------***---------------------------------------------------------------------------surveillance applications, AMC must operate reliably Abstract-Automatic modulation classification (AMC) without prior knowledge of transmission parameters and methods trained on synthetic datasets and end-to-end under nonideal channel and hardware conditions [2]. deep learning models often exhibit limited robustness in However, contemporary ML and DL based AMC methods, noisy real-world RF environments and provide little although effective on synthetic IQ benchmarks, often suffer interpretability for operational use. This paper proposes a performance degradation in real-world RF settings due to human-centered hybrid software defined radio (SDR) noise, frequency offsets, fading, and receiver impairments intelligence framework that combines dig-ital signal [3]. A further limitation is the limited interpretability of processing (DSP), machine learning (ML), and a neural classifiers, which constrains trust, traceability, and governance-aware explainable artificial intelligence (XAI) operational accountability in mission-critical deployments layer for interpretable modulation classification. Real[4], [5]. time IQ captures are acquired using HackRF One hardware and GNU Radio pipelines under practical SDR To address these issues, this paper proposes a humanoperating conditions. From each capture, a 14centered hybrid SDR intelligence framework that dimensional feature vector is extracted using FFT-based combines DSP, ML, and governance-aware XAI for spectral analysis, occupied bandwidth estimation, signal interpretable mod-ulation classification on real SDR power analysis, instantaneous frequency, and captures. The framework extracts 14 transparent RF constellation anal-ysis, and is then classified by a Random features from IQ data acquired using HackRF One and GNU Forest model. The framework classifies FSK-like digital Radio, and employs confidence-based decision fusion signals, Narrowband FM (NBFM), Wideband FM (WBFM), together with human-in-the-loop over-sight to improve and no-signal conditions with 100% accuracy, precision, decision transparency and operational usabil-ity. recall, and F1-score on the evaluated real-SDR dataset. A Experimental results on FSK-like digital signals, NBFM, human-in-the-loop governance architecture further WBFM, and no-signal captures demonstrate the provides DSP-output visibility, confidence-based decision feasibility of lightweight and explainable AMC for fusion, ethical RF monitoring, and audit logging. These controlled real-SDR scenarios. results indicate that the proposed framework is a lightweight and interpretable solution for controlled SDRThe main contributions are summarized as follows: based RF monitoring scenarios. Index Terms-Software Defined Radio (SDR), IQ Signal Pro-cessing, Automatic Modulation Classification (AMC), Explainable Artificial Intelligence (XAI), Digital Signal Processing (DSP), RF Intelligence, Machine Learning (ML), Tactical Communication Systems.
I. INTRODUCTION Software defined radio (SDR) enables flexible, softwaredriven adaptation of modulation, filtering, spectrum monitor-ing, and waveform analysis, making it highly relevant for automatic modulation classification (AMC) in tactical RF en-vironments [1]. In military communication, electronic warfare, and spectrum
© 2026, IRJET
|
Impact Factor value: 8.315
|
A hybrid DSP–ML AMC framework implemented on real HackRF One/GNU Radio IQ captures. • A 14-dimensional interpretable RF feature set derived from explicit DSP formulations. • A governance-aware human-centered architecture incor-porating confidence fusion, operator visibility, and audit logging. • Experimental validation on real SDR captures for four signal classes, achieving 100% accuracy on the evaluated dataset. •
The remainder of this paper is organized as follows. Section II reviews related work, followed by the proposed system architecture, feature extraction, classification methodology, experiments, and conclusions.
ISO 9001:2008 Certified Journal
|
Page 235