PCA-Based Adversarial Attacks on Signature Verification Systems

Handwritten Signatures are popular biometrics that are used to authenticate individuals based on their unique physical or behavioral attributes. These systems rely on deep neural networks (DNN) for feature extraction. However, DNNs are vulnerable to small imperceptible perturbations. This research evaluates the robustness of signature verification systems through adversarial attacks. The state-of-the-art adversarial attacks possess certain limitations: first attacks being white-box in nature are impractical, and second these attacks add noise to the whole image including the background making them quite perceptible. To address these limitations, this study presents a lightweight approach based on principal component analysis (PCA). Our novel algorithm generates a universal noise vector using spatial transformation on principal components. It also strategically confines perturbation to specific regions while exploiting the principal components of the input image to launch attacks. We also computed evaluations conducted across three benchmark datasets to demonstrate compelling outcomes, in terms of attack success rate, imperceptibility, and transferability.