Traditional methods for classifying Global Navigation Satellite System (GNSS) jamming signals heavily rely on post-processing raw or spectral data streams, which demands complex, high-bandwidth, and costly data transmission to cloud-based systems. To address this challenge, a novel research paper presented at the IEEE/ION PLANS 2025 symposium introduces an edge-centric solution: compressing GNSS data streams directly at the hardware receiver while simultaneously classifying jamming and spoofing attacks in real time.
This hardware-focused approach deploys Generative AI (GenAI), specifically Variational Autoencoders (VAEs), on Google Edge Tensor Processing Units (TPUs). The study evaluates various autoencoder (AE) architectures to compress and reconstruct GNSS signals, prioritizing the preservation of critical interference characteristics while minimizing the data footprint near the receiver. To match the strict power constraints of edge environments, the pipeline adapts large-scale AE models for Google Edge TPUs using 8-bit quantization, guaranteeing energy-efficient execution.
Evaluated against raw in-phase and quadrature-phase (IQ) data, Fast Fourier Transform (FFT) data, and handcrafted features, the system achieved a remarkable data compression ratio of over 42x. More importantly, the classification of approximately 72 interference types on the reconstructed signals yielded an F2-score of 0.915, closely matching the 0.923 F2-score achieved on original, uncompressed signals. This edge-native pipeline substantially reduces the data transmission overhead, offering a highly practical solution for real-time interference mitigation.
To bolster user trust in sensitive security applications, the researchers conducted ablation studies on conditional and factorized VAEs (e.g., FactorVAE). By exploring latent feature disentanglement for data generation, they enhanced model interpretability, shedding light on how neural networks differentiate and represent complex signal threats in resource-constrained environments.
[AgentUpdate Depth Analysis] This research marks a pivotal milestone for the deployment of autonomous edge AI Agents in physical-world environments. Typically, multimodal agents are constrained by the sheer volume of raw sensor data they must process or transmit. By proving that Generative AI—specifically FactorVAEs—can run locally on ultra-low-power accelerators like the Google Edge TPU to achieve 42x compression and near-lossless feature preservation, this work establishes a blueprint for future "sensor-to-agent" architectures. Autonomous agents operating in disrupted or communication-denied environments (such as tactical drones or maritime vessels) can leverage similar localized VAEs to compress ambient RF or telemetry streams into high-fidelity, interpretable latent spaces. This turns raw, high-dimensional physical signals into semantic representations that local reasoning engines can parse in real-time, enabling resilient, low-latency, and decentralized agent decisions without cloud reliance.
