SpectrumFM: A Foundation Model for Intelligent Spectrum Management

Intelligent spectrum management is crucial for improving spectrum efficiency and achieving secure utilization of spectrum resources. However, existing intelligent spectrum management methods, typically based on small-scale models, suffer from notable limitations in recognition accuracy, convergence speed, and generalization, particularly in the complex and dynamic spectrum environments. To address these challenges, this paper proposes a novel spectrum foundation model, termed SpectrumFM, establishing a new paradigm for spectrum management. SpectrumFM features an innovative encoder architecture that synergistically exploits the convolutional neural networks and the multi-head self-attention mechanisms to enhance feature extraction and enable robust representation learning. The model is pre-trained via two novel self-supervised learning tasks, namely masked reconstruction and next-slot signal prediction, which leverage large-scale in-phase and quadrature (IQ) data to achieve comprehensive and transferable spectrum representations. Furthermore, a parameter-efficient fine-tuning strategy is proposed to enable SpectrumFM to adapt to various downstream spectrum management tasks, including automatic modulation classification (AMC), wireless technology classification (WTC), spectrum sensing (SS), and anomaly detection (AD). Extensive experiments demonstrate that SpectrumFM achieves superior performance in terms of accuracy, robustness, adaptability, few-shot learning efficiency, and convergence speed, consistently outperforming conventional methods across multiple benchmarks. Specifically, SpectrumFM improves AMC accuracy by up to 12.1% and WTC accuracy by 9.3%, achieves an area under the curve (AUC) of 0.97 in SS at -4 dB signal-to-noise ratio (SNR), and enhances AD performance by over 10%.

💡 Research Summary

The paper introduces SpectrumFM, a foundation model specifically designed for intelligent spectrum management across a variety of tasks such as automatic modulation classification (AMC), wireless technology classification (WTC), spectrum sensing (SS), and anomaly detection (AD). Recognizing the limitations of existing small‑scale, task‑specific models—namely low recognition accuracy, slow convergence, and poor generalization in dynamic RF environments—the authors propose a large‑scale, self‑supervised approach that learns universal spectrum representations from massive in‑phase and quadrature (IQ) datasets.

The core of SpectrumFM is a hybrid encoder that combines convolutional neural networks (CNNs) with multi‑head self‑attention (MHSA). The CNN layers capture fine‑grained, localized spectral patterns, while the MHSA modules model long‑range temporal and inter‑frequency dependencies, yielding robust, high‑level feature embeddings. Two novel self‑supervised pre‑training tasks are employed: (1) masked reconstruction, where random portions of the IQ sequence are masked and the model must reconstruct them, encouraging the learning of latent structure and resilience to noise; and (2) next‑slot signal prediction, which forces the model to forecast the IQ values of the subsequent time slot, thereby endowing it with temporal awareness essential for dynamic spectrum scenarios. Both tasks require no labeled data, allowing the model to be pre‑trained on millions of unlabeled samples.

To adapt the pre‑trained model to downstream tasks, the authors adopt a parameter‑efficient fine‑tuning strategy based on low‑rank adapters (e.g., LoRA). Only a small subset of parameters is updated, enabling rapid convergence and strong few‑shot performance without the need for large labeled datasets for each task.

Extensive experiments demonstrate that SpectrumFM consistently outperforms state‑of‑the‑art baselines. In AMC, it improves accuracy by up to 12.1 percentage points on the RML2016.04C dataset and yields gains of 7.5 pp and 1.9 pp on the RML2016.10A/B datasets. For WTC, it achieves a 9.3 pp increase over the best existing method. In spectrum sensing, the model attains an AUC of 0.97 even at a challenging –4 dB SNR, surpassing traditional energy‑detector limits. In anomaly detection, it raises AUC by more than 10 pp, indicating superior capability to identify illicit or abnormal spectrum usage. Moreover, convergence speed improves by roughly 30 % compared with conventional architectures, highlighting its suitability for real‑time deployment.

The authors discuss several strengths: the hybrid encoder balances spatial detail and global context; the self‑supervised tasks foster robustness to noise, interference, and non‑stationary conditions; and the fine‑tuning scheme enables efficient transfer to new tasks with minimal data. Limitations are also acknowledged: the current evaluation focuses on sub‑GHz bands, leaving mmWave and THz regimes untested; the pre‑training data are primarily laboratory‑collected IQ samples, so real‑world hardware non‑linearities and channel uncertainties may require additional adaptation.

Future work suggested includes extending the model to multi‑modal inputs (e.g., spectrograms, power spectral density, geographic metadata), incorporating online continual learning to handle evolving spectrum environments, and exploring model compression techniques for edge deployment.

In summary, SpectrumFM establishes a new paradigm for spectrum management by providing a versatile, high‑performance foundation model that can be efficiently adapted to a wide range of RF tasks, thereby advancing both spectrum efficiency and security in next‑generation wireless networks.