MiguelSantPasc/Code_AdvSSM: v1.0.0 – Adversarial Observations in State-Space Models for Robust RL Copiar

Abstract

Code release – v1.0.0 This release contains the source code accompanying the manuscript "Adversarial Observations in State-Space Models for Robust Reinforcement Learning" by M. Santos-Pascual and D. Ríos Insua. The repository provides implementations and experimental scripts for studying adversarial perturbations of observations in probabilistic state-space models and their impact on downstream reinforcement learning. The main experimental components include: Adversarial observation attacks against Kalman filtering and smoothing. Evaluation of adversarial attacks in reinforcement learning environments, including navigation and CartPole. Analysis of the influence of observation and process noise covariance on attack effectiveness. Evaluation of uncertainty-aware defensive strategies based on adaptive covariance inflation. Comparison of attack and defense performance under different experimental settings. This version is intended to provide a reproducible snapshot of the computational experiments accompanying the manuscript. Archival information: This release is archived through Zenodo to provide a persistent DOI for research reproducibility and citation.