GeoShield: An Adversarially Robust Defense Framework Against Map Tampering in Autonomous Navigation

Authors:

Prof. Roopesh Kumar B N1, Vijayalakshmi C2, Talluru Sahithya 3, Shrihari I B4, Shiva M5 1Associate Professor, Department of CSE, K.S. Institute of Technology (KSIT), Bengaluru, India

2Student, Department of CSE, K.S. Institute of Technology (KSIT), Bengaluru, India. 3Student, Department of CSE, K.S. Institute of Technology (KSIT), Bengaluru, India. 4Student, Department of CSE, K.S. Institute of Technology (KSIT), Bengaluru, India. 5Student, Department of CSE, K.S. Institute of Technology (KSIT), Bengaluru, India.

Email: 1roopeshkumarbn@ksit.edu.in,2vijayalakshmics05@gmail.com, 3sahithya6360@gmail.com, 4ibshrihari@gmail.com, 5sm3467928@gmail.com

Abstract- High-definition (HD) maps and geolocation metadata play a crucial role in autonomous vehicle navigation, but they create new cybersecurity risks. This paper introduces GeoShield, a multi-layered defense framework that detects and corrects adversarial map tampering attacks using OpenStreetMap (OSM) metadata alongside NASA SRTM elevation data. GeoShield simulates realistic attack scenarios—tunnel/bridge label flipping, road-direction manipulation, and speed-limit spoofing. It uses fast heuristic validation, including elevation consistency and road connectivity checks, combined with graph-based anomaly detection models like Isolation Forest, Hidden Markov Models (HMMs), and GraphSAGE. When it identifies suspicious road segments, GeoShield revokes trust and performs selective re-verification with the Overpass API, regaining reliable metadata to restore accurate information. When deployed on European and U.S. road networks, Tier-1 and Tier-2 detection modules are designed to flag anomalous map segments, while Tier-3 verification is expected to confirm tampering with high accuracy. .Geospatial metadata validation can function as an effective cybersecurity barrier, strengthening the robustness and safety of map-driven ADAS systems.

Keywords- Autonomous navigation, adversarial attacks, OpenStreetMap, graph neural networks, anomaly detection