GAC-KAN: An Ultra-Lightweight GNSS Interference Classifier for GenAI-Powered Consumer Edge Devices

The integration of Generative AI (GenAI) into Consumer Electronics (CE)–from AI-powered assistants in wearables to generative planning in autonomous Uncrewed Aerial Vehicles (UAVs)–has revolutionized user experiences. However, these GenAI applications impose immense computational burdens on edge hardware, leaving strictly limited resources for fundamental security tasks like Global Navigation Satellite System (GNSS) signal protection. Furthermore, training robust classifiers for such devices is hindered by the scarcity of real-world interference data. To address the dual challenges of data scarcity and the extreme efficiency required by the GenAI era, this paper proposes a novel framework named GAC-KAN. First, we adopt a physics-guided simulation approach to synthesize a large-scale, high-fidelity jamming dataset, mitigating the data bottleneck. Second, to reconcile high accuracy with the stringent resource constraints of GenAI-native chips, we design a Multi-Scale Ghost-ACB-Coordinate (MS-GAC) backbone. This backbone combines Asymmetric Convolution Blocks (ACB) and Ghost modules to extract rich spectral-temporal features with minimal redundancy. Replacing the traditional Multi-Layer Perceptron (MLP) decision head, we introduce a Kolmogorov-Arnold Network (KAN), which employs learnable spline activation functions to achieve superior non-linear mapping capabilities with significantly fewer parameters. Experimental results demonstrate that GAC-KAN achieves an overall accuracy of 98.0%, outperforming state-of-the-art baselines. Significantly, the model contains only 0.13 million parameter–approximately 660 times fewer than Vision Transformer (ViT) baselines. This extreme lightweight characteristic makes GAC-KAN an ideal “always-on” security companion, ensuring GNSS reliability without contending for the computational resources required by primary GenAI tasks.

💡 Research Summary

The paper introduces GAC‑KAN, an ultra‑lightweight deep‑learning framework designed to detect GNSS jamming on consumer edge devices that also run generative AI (GenAI) workloads. The authors first address the chronic shortage of real‑world interference data by constructing a physics‑guided simulation pipeline. Six fundamental jamming primitives—single‑tone, multi‑tone, linear frequency‑modulated (LFM) chirp, pulse, partial‑band noise (PBNJ), and sinusoidal‑chirp interference (SCI)—are mathematically modeled with randomized parameters (carrier frequency, power, phase, sweep rate, duty cycle, etc.). These synthetic signals are mixed with additive white Gaussian noise and processed through a Short‑Time Fourier Transform (STFT) to produce 2‑D spectrograms that faithfully capture both time and frequency characteristics of each interference type.

The network architecture consists of two main components. The feature‑extraction backbone, called Multi‑Scale Ghost‑ACB‑Coordinate (MS‑GAC), merges GhostNet’s cheap linear operations with Asymmetric Convolution Blocks (ACB). ACB splits a conventional 3×3 kernel into separate horizontal and vertical convolutions during training, enriching the receptive field without increasing inference cost. Ghost modules then generate redundant feature maps via inexpensive linear transformations, drastically reducing parameter count. The backbone processes spectrograms at multiple spatial scales, allowing coarse‑level global frequency patterns and fine‑level transient details (e.g., pulses) to be captured simultaneously. A coordinate‑attention mechanism re‑encodes positional information along the time and frequency axes, ensuring the network focuses on the most informative regions of the spectrogram.

Instead of a conventional fully‑connected classifier, the authors replace it with a Kolmogorov‑Arnold Network (KAN). KAN places learnable spline activation functions on each edge of the linear layer, providing highly expressive non‑linear mappings while using far fewer parameters than a standard MLP. This design is especially advantageous for distinguishing non‑linear interference such as SCI, where traditional ReLU‑based heads struggle.

Experimental evaluation compares GAC‑KAN against Vision Transformer (ViT), ResNet, MobileNet, and other state‑of‑the‑art lightweight models. GAC‑KAN achieves 98.0 % overall classification accuracy on the synthetic dataset while containing only 0.13 M parameters—approximately 660× fewer than ViT baselines. FLOP counts and memory footprints are low enough for real‑time, always‑on inference on typical consumer SoCs that already host GenAI models. Ablation studies show that the combination of multi‑scale Ghost‑ACB features and the KAN head contributes the most to performance gains, with KAN alone improving accuracy by 2–3 % over an equivalent MLP head, particularly on the SCI class.

In conclusion, GAC‑KAN simultaneously solves the data‑scarcity problem through high‑fidelity physics‑based simulation and the extreme efficiency problem through a novel backbone plus spline‑based classifier. Its 0.13 M‑parameter footprint makes it suitable for deployment as a background security monitor that does not compete with primary GenAI workloads, thereby preserving GNSS reliability in next‑generation consumer electronics. Future work may involve validating the model on over‑the‑air captured jamming signals and integrating hardware‑aware quantization for further power savings.