Dualformer: Efficient Feature Extractor for Complex-valued Blind Communication Signal Analysis

Technical Summary: Dualformer: Efficient Feature Extractor for Complex-valued Blind Communication Signal Analysis

Problem Statement
Blind communication signal analysis, encompassing Automatic Modulation Recognition (AMR), Signal Scheme Recognition (SSR), and Signal Structure Parsing (SSP), relies heavily on extracting effective features from complex-valued In-phase and Quadrature (IQ) sequences. Current Deep Learning (DL) approaches generally fall into two paradigms: Real-Valued Neural Networks (RVNNs), which process I and Q components separately or concatenated, and Complex-Valued Neural Networks (CVNNs), which process complex numbers natively.
RVNNs often lack a rigorous theoretical foundation for exploiting the intrinsic complex structure and may fail to capture phase relationships effectively. Conversely, while CVNNs offer enhanced representational capacity, they suffer from increased computational complexity, optimization instability due to complex error surfaces, and a higher demand for training data to avoid overfitting. The core challenge lies in balancing approximation error (model capacity) and estimation error (generalization) while efficiently leveraging the statistical symmetry of IQ channels without incurring the optimization costs of full CVNNs.

Methodology
The authors propose DualNN, a dual-channel neural network framework that processes the real and imaginary parts of complex-valued signals using shared network parameters. Unlike CVNNs that use complex weights, DualNN employs a Real-Valued Neural Network (RVNN) backbone where the same weights are applied to both the $x_{real}$ and $x_{imag}$ inputs.

• Theoretical Basis: The authors argue that many communication signals (e.g., QAM, PSK) possess circular symmetry, implying that I and Q channels share identical second-order statistics. By sharing parameters, DualNN explicitly leverages this physical symmetry. Theoretical analysis based on generalization error bounds demonstrates that DualNN achieves a tighter generalization error than both RVNNs and CVNNs. Specifically, it reduces the estimation error (by stabilizing the optimization landscape and restricting the spectral norm of weight matrices) while maintaining approximation capabilities comparable to CVNNs, particularly in data-scarce scenarios.
• Architecture (Dualformer): To instantiate DualNN, the authors propose Dualformer, a Transformer-based architecture.
  • Tokenization: Input signals are segmented into patches to reduce computational complexity and capture long-range dependencies.
  • Dual-Channel Processing: The real and imaginary patch sequences are fed into a shared Transformer encoder.
  • Projection Headers: The architecture includes task-specific heads:
    • Header V1: A coarse-grained header (2-layer CNN + pooling) for AMR and SSR tasks to extract global semantic features.
    • Header V2: A fine-grained header (single-layer CNN, no pooling) for SSP tasks to preserve temporal resolution for point-by-point segmentation.

Key Contributions

1. DualNN Framework: Introduction of a parameter-sharing architecture for complex-valued signals that stabilizes training and reduces model complexity compared to CVNNs.
2. Theoretical Guarantees: Mathematical proofs rooted in generalization theory showing that DualNN achieves a tighter generalization error bound than RVNNs universally and outperforms CVNNs under limited training data by balancing approximation and estimation errors.
3. Dualformer Implementation: A novel Transformer-based module with specialized projection headers (V1 and V2) designed to handle both coarse-grained (AMR/SSR) and fine-grained (SSP) blind signal analysis tasks.
4. Empirical Validation: Extensive experiments demonstrating consistent performance improvements over state-of-the-art Transformer variants and conventional DL models across AMR, SSR, and SSP tasks.

Experimental Results
The paper evaluates Dualformer against three Transformer-based baselines (Transformer, Informer, Autoformer) and four conventional DL approaches (CNN, LSTM, ResNet, etc.) across three datasets: a synthetic AMR dataset, the RML 2018.01A benchmark, and a custom SSR/SSP dataset.

• Performance: Dualformer consistently outperforms all baselines in accuracy and F1-score across varying Signal-to-Noise Ratios (SNR) from -4 dB to 12 dB. Notably, it shows superior robustness at low SNR (-4 dB) compared to CVNNs, attributed to the self-attention mechanism's ability to aggregate long-range phase dependencies.
• Efficiency and Data Scarcity: In sample efficiency tests, DualNN demonstrated stable accuracy with significantly fewer training samples compared to CVNNs, which struggled in low-data regimes.
• Robustness: Under I/Q imbalance conditions (simulating hardware imperfections), DualNN maintained performance close to CVNNs (within 0.10% gap) while significantly outperforming RVNNs, confirming that the parameter-sharing design does not incur a meaningful robustness cost.
• Complexity: Despite having roughly half the parameters and FLOPs of a comparable CVNN, DualNN achieved competitive or superior accuracy, validating the efficiency of the shared-parameter approach.

Significance
The paper posits that DualNN provides a new paradigm for processing complex-valued signals in blind analysis, effectively bridging the gap between the representational power of CVNNs and the optimization stability of RVNNs. By theoretically and empirically demonstrating that parameter sharing reduces generalization error, the work suggests that explicit complex-valued operations are not strictly necessary for high performance if the underlying statistical symmetries are properly exploited. The authors claim this approach paves the way for broader applications in unsupervised and weakly supervised scenarios, such as blind source separation and low-SNR spectrum sensing, where data scarcity and model stability are critical constraints.