New Adversarial Attack Targets Online Handwriting Recognition

Deep learning models used for online handwriting recognition are vulnerable to adversarial attacks, a challenge that existing image-based attack methods struggle with due to their reliance on spatial perturbations. These traditional methods often create unnatural artifacts when applied to the time-series nature of handwriting data. This new research introduces a framework that generates adversarial examples by subtly editing the temporal sequence of pen strokes. The proposed technique identifies "salient" time steps in the handwriting trajectory using gradient-based activation mapping. At these critical points, the method inserts or deletes data points, effectively altering the input without introducing visible jitter or unnatural shapes. This temporal editing approach demonstrates superior transferability in black-box attack scenarios compared to conventional image-based attacks, highlighting a significant new threat model for online handwriting systems.