Title: Compromising Robot Localization Through Deep Feature Manipulation

Robot localization is a cornerstone of autonomous navigation and operational safety. However, these systems are susceptible to adversarial perturbations that can induce mislocalization, navigation failures, or hazardous interactions, particularly in high-stakes missions. This study examines the susceptibility of deep learning-driven localization pipelines to such attacks. We introduce a novel framework designed to generate adversarial queries that specifically exploit Product Quantization (PQ) within visual localization systems.

The proposed method utilizes a Lightweight Product Quantization Network (LPQN) to alter query feature encodings. By doing so, it confuses the retrieval mechanism, causing the system to return database entries that are semantically unrelated to the actual location. The generation of these adversarial queries follows a two-step process: a forward pass that disrupts feature distributions, followed by a backward pass that fine-tunes the perturbation via optimization. Thanks to the efficient architecture of the LPQN, it is possible to produce subtle yet potent perturbations with negligible computational cost. Comprehensive testing in both controlled settings and real-world robotic environments reveals that our approach significantly impairs PQN performance, highlighting significant security flaws in practical deployments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC