<p>Metal corrosion is a major degradation mechanism that reduces the load-carrying capacity and serviceability of steel structural members. Quantitative identification of corrosion-induced stiffness loss from limited structural response measurements remains a challenging inverse problem, particularly under sparse sensing and measurement noise. This study presents a physics-informed neural network (PINN)-based framework for identifying spatially varying flexural rigidity in steel I-beams. The proposed approach embeds the Euler–Bernoulli beam equation directly into the loss function, enabling simultaneous inference of the transverse displacement field and the stiffness degradation field from sparse displacement data. The framework is evaluated through numerical case studies involving uniform, localized, and multi-point corrosion scenarios. Gaussian noise levels of up to 5% are considered to examine robustness under measurement uncertainty. The results indicate that the proposed method can identify the location and magnitude of stiffness degradation with low estimation error under the investigated conditions. Comparisons with a conventional finite element model updating approach and a purely data-driven deep neural network model show improved identification accuracy and localization capability for the examined benchmark problem. The findings suggest that embedding governing mechanical equations into neural network training enhances solution stability and data efficiency in inverse stiffness identification problems. Nevertheless, the present validation is limited to numerical simulations, and further experimental investigation is required to assess field applicability.</p>