A Speculative GLRT-Backed ApproachRobust Deep Learning-Based Array Processing

Technical Summary: A Speculative GLRT-Backed Approach for Robust Deep Learning-Based Array Processing

Problem Statement
Deep learning (DL) has emerged as a low-latency alternative to classical statistical methods for array processing tasks, specifically signal detection and Direction of Arrival (DoA) estimation. However, DL models suffer from a lack of statistical interpretability and, critically, are highly susceptible to adversarial interference. In wireless environments, adversaries can introduce subtle, over-the-air additive perturbations to transmitted signals to induce misclassification at the receiver. While traditional likelihood-based methods (e.g., Maximum Likelihood Estimation, Generalized Likelihood Ratio Test) offer statistical rigor and robustness, they are computationally intensive, resulting in high latency unsuitable for real-time applications. Conversely, standalone DL models lack robustness, and existing defenses (such as adversarial training or denoising) often fail to provide significant performance improvements while ignoring the inherent statistical properties of array processing.

Methodology
The authors propose a "speculative array processing framework" that synergizes the speed of DL with the statistical reliability of the Generalized Likelihood Ratio Test (GLRT). The core of this approach relies on a theoretical finding regarding the spatial robustness of second-order statistics.

1. Theoretical Foundation: The paper establishes that the spatial domain (specifically the covariance matrix) of a received array signal is robust to $\ell_p$-bounded adversarial perturbations. The authors prove (Theorem 1) that for perturbations $\delta$ where $\|\delta\|_p \leq \epsilon$, the deviation in the covariance of the received signal is negligible relative to the signal energy. This implies that while time-domain perturbations can fool DL classifiers, their impact is averaged out in the spatial domain, preserving the integrity of GLRT estimations.

2. Speculative Framework (Algorithm 1):

   • Speculative Stage (DL): A low-latency Deep Learning classifier (Convolutional Neural Network) processes the raw received data (for detection) or empirical covariance matrices (for DoA) to generate an immediate, "speculative" prediction. This allows the receiver to proceed with post-processing tasks (e.g., decoding) without waiting for slower statistical computations.
   • Validation Stage (GLRT): Simultaneously, a GLRT estimator operates on the same signal. For signal detection, it compares the trace of the product of empirical covariance matrices from consecutive time windows. For DoA estimation, it utilizes a matched filter approach on the covariance matrices to estimate the angle.
   • Consistency Check: Once the GLRT inference is complete (which takes longer due to computational cost), a consistency check is performed. If the DL prediction matches the GLRT result, the speculative decision is confirmed. If they disagree, the system discards the DL output and reverts to the GLRT result, ensuring robustness at the cost of a potential re-processing delay.

3. Adversarial Setup: The framework is evaluated against Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks under both $\ell_2$ and $\ell_\infty$ norms, measured by the Perturbation-to-Signal Ratio (PSR).

Key Contributions

• Spatial Robustness Proof: The paper theoretically demonstrates that $\ell_p$-bounded adversarial interference has a negligible effect on the second-order statistics (covariance) of received array signals, thereby mitigating attack potency in the spatial domain.
• Hybrid Framework: The authors introduce the first framework to jointly utilize a DL-MLE architecture for adversarial robustness in array processing. This speculative approach integrates a fast DL classifier with a theoretically grounded GLRT validator.
• Empirical Validation: The study provides empirical evidence that this hybrid approach outperforms state-of-the-art baselines, including adversarial training, defensive distillation, and denoising autoencoders (DAE), across various perturbation magnitudes and designs.

Results
Experimental evaluations were conducted using a Uniform Linear Array (ULA) with 8 sensors for both binary signal detection and multi-class DoA estimation (61 classes).

• Baseline Performance: Standard CNNs showed a sharp decline in accuracy as the PSR increased. Adversarial training provided partial recovery, and DAE pre-processing showed limited success at moderate PSRs but collapsed at higher perturbation levels.
• GLRT Robustness: The GLRT component retained high accuracy across all tested PSR ranges and attack types, validating the theoretical claim of spatial robustness.
• Speculative Framework Performance: In non-adversarial conditions (low PSR), the speculative DL classifier achieved performance equivalent to the GLRT, confirming its reliability for fast inference. Under adversarial conditions, the framework successfully leveraged the GLRT to correct DL misclassifications, demonstrating superior robustness compared to standalone DL and other defense mechanisms.

Significance and Claims
The paper claims to address a critical research gap by providing a solution that offers both low-latency inference and adversarial robustness with statistical interpretability. The authors position this work as the first framework to successfully combine DL and GLRT architectures specifically for adversarial robustness in array processing. The significance lies in the ability to maintain real-time processing speeds (via DL speculation) while ensuring decision reliability through a theoretically grounded statistical validator (GLRT), effectively neutralizing the vulnerability of DL models to over-the-air adversarial attacks without sacrificing the interpretability of classical signal processing. Future work is noted to potentially extend this framework to other DL-based vulnerabilities such as Multiple Signal Classification (MUSIC), Automatic Modulation Classification (AMC), and MIMO systems.