STOCHASTIC-SPECTRAL METHODS FOR BACKDOOR DETECTION IN NEURAL NETWORKS FOR 6G SIGNAL PROCESSING

Abstract

Modern 6G communication systems utilize deep neural networks (DNNs) for adaptive signal processing and spectral resource optimization. Simultaneously, the risk of backdoor attacks is increasing, where hidden triggers embedded in the network's weights can intentionally alter its behavior – a critical vulnerability for the physical layer of high-frequency signals. This study proposes mathematically rigorous methods for detecting such hidden triggers, based on a combination of stochastic signal modeling, spectral analysis, and information-theoretic entropy metrics.

6G signals are formalized as multivariate stochastic processes with known statistical properties, enabling the formalization of anomalous modifications introduced by hidden triggers. Neural networks are treated as non-linear operators that transform signals into the output space, providing a mathematical definition of the network's sensitivity to potential attacks. The methodology integrates spectral and wavelet signal analysis, singular value decomposition (SVD) of weight matrices, and an assessment of information interconnectivity, which enables the detection of anomalies within both the network structure and the signal spectrum.

A unified detection criterion is proposed, combining spectral, stochastic, and information-theoretic features to ensure the localization of anomalous components and a formalized assessment of network resilience. Experimental validation involves 6G signal simulation and testing across various neural network architectures, allowing for an evaluation of the accuracy, sensitivity, and reliability of the proposed methods. The expected outcome is a mathematical framework for the secure operation of neural networks at the 6G physical layer, which enhances the resilience of high-frequency communications against hidden triggers and establishes a scientific basis for the further development of detection algorithms and security assessment of signals in complex telecommunication systems.