Privacy Concerns of DAS: Eavesdropping using Neural Network Transcription

Distributed acoustic sensing (DAS) has transformed geophysical, environmental, and infrastructure monitoring. However, the increasing bandwidth and sensitivity of modern interrogators now extend into the audio range, introducing a material privacy risk. Here we demonstrate, through in-situ experiments on live fibre deployments, that human speech, music, and other acoustic signals can be under certain acquisition conditions.

We show that intelligible speech can be accurately recovered and automatically transcribed using neural networks. Experiments were conducted on both linear and spooled fibre geometries, deployed as part of an ongoing geophysical survey. We find that coiled layouts, which are common in access networks (e.g., slack loops or storage spools), exhibit enhanced sensitivity to incident acoustic waves relative to linear layouts. Modelling indicates this arises from increased broadside sensitivity and reduced destructive interference for longer wavelength acoustic fields over the gauge length. We systematically assess how acquisition parameters, such as source-fibre offset, influence signal‑to‑noise ratio, spectral fidelity, and speech intelligibility of recorded audio. We further show that neural network based denoising strategies improves intelligibility and fidelity of recorded audio, thereby exacerbating privacy concerns.

These findings demonstrate that appropriate interrogation of existing fibre infrastructure - including fibre‑to‑the‑premises links, smart-city infrastructure, and research cables – can function as pervasive, passive wide-area acoustic receivers, creating a pathway for inadvertent or malicious eavesdropping. We discuss practical mitigation strategies spanning survey design, interrogation configuration, and data governance, and argue that the incorporation of privacy‑by‑design into deployment and processing is crucial to leverage the unique benefits of DAS while managing emerging ethical and legal risks.