Abstract
Handwritten signature verification has increasingly adopted contrastive learning frameworks, such as Siamese networks, to distinguish between genuine and forged samples. While robust against random noise, these models exhibit a critical vulnerability to targeted adversarial attacks. This paper presents a novel white-box False Positive (FP) attack that frames adversarial generation as a style transfer problem. By optimizing a synthesized image to match the stylistic features of a genuine signature while retaining the content of a forgery, the proposed method achieves success rates up to 99.71%, far exceeding traditional gradient-based attacks.
Introduction
Handwritten signature verification is a vital biometric authentication method. Modern systems utilize contrastive learning to map signatures into a latent space where genuine pairs are clustered together and forgeries are pushed apart. However, the sparse nature of signatures—thin lines on clean backgrounds—makes them uniquely challenging for standard adversarial attacks like FGSM or PGD. These traditional methods often introduce perceptible noise that contrastive models easily ignore or detect, failing to move the forgery's embedding close enough to the target genuine signature. This study addresses the False Positive attack: modifying a forgery so it is incorrectly classified as genuine.
Methodology
The proposed attack treats the genuine signature as a "style" source and the forgery as "content." The objective is to synthesize an image that retains the structure of the forgery but adopts the stylistic nuances (stroke width, pressure, curvature) of the genuine writer.
The optimization process employs the Adam optimizer to update the pixels of a synthesized image, initialized from the forged image, while keeping the target model's weights frozen. The optimization is guided by a composite loss function consisting of four parts: Attack Loss (latent space alignment), Style Loss (Gram matrix alignment), Difference Loss (pixel-level stealth), and Total Variation Loss (smoothness).
Key Results
The method was evaluated on the CEDAR (English) and BHSig260-B (Bengali) datasets. The results demonstrate a significant performance gap between the proposed style-aware attack and traditional methods.
Performance Comparison
On the CEDAR dataset, the proposed method achieved a near-perfect 99.71% success rate. In contrast, standard attacks like PGD often resulted in a 0% success rate for False Positive attacks on these specific contrastive architectures. On the more linguistically complex BHSig260-B dataset, the success rate reached 82.93%. Visually, the proposed attack maintains the integrity of the signature far better than traditional noise-based methods, which often produce heavily distorted images.
Ablation and Robustness
An ablation study highlighted the necessity of the Style Loss component. Removing style alignment caused the success rate on BHSig260-B to drop from 82.93% to 65.02%. Furthermore, the attack proved robust against common defenses; even when the target model underwent adversarial training, the attack maintained a 93.7% success rate on CEDAR, suggesting that current defenses are insufficient against style-aware perturbations.
Discussion and Conclusion
The success of this method stems from its ability to "re-style" the forgery rather than merely adding noise. By focusing perturbations on the foreground (the ink) and aligning high-level stylistic features, the attack effectively bypasses the sensitivity of contrastive models to local feature disruptions.
While highly effective in a white-box setting, future work must address the transition to black-box scenarios where model parameters are unknown. Additionally, the requirement for a target genuine signature for style extraction remains a practical constraint. This research underscores the need for more sophisticated defense mechanisms that can detect subtle stylistic perturbations in biometric verification systems.
Summary updated