Adversarial Robustness Evaluation of a Vietnamese Handwriting OCR System

Deep learning has revolutionized optical character recognition (OCR), enabling significant improvements in handwritten text recognition. However, the vulnerability of these models to adversarial manipulation poses a challenge, particularly in applications where robustness is essential. This paper examines the susceptibility of a Vietnamese handwriting OCR model to white-box adversarial attacks. Using the projected gradient descent (PGD) method, we craft imperceptible perturbations that manipulate the model predictions while preserving the visual integrity of the inputs. Our empirical study analyzes the impact of adversarial perturbations on recognition accuracy, highlighting the fragility of advanced OCR systems against such attacks. Source code is available at: https://github.com/ELO-Lab/AdvRE-ViHandOCR.