Deep learning models, despite their impressive performance, remain vulnerable to adversarial attacks. Subtle perturbations that can significantly disrupt model predictions. Generating adversarial examples with these imperceptible alterations to both global and salient features often demands substantial computational resources. In this paper, we propose an enhancement to existing adversarial attack methods by incorporating a non-gradient-based approach using vector quantization (VQ) as a preprocessing step. By leveraging VQ’s encoding capability, our method optimizes the codebook used for attacks, strengthening the performance of both projected gradient descent (PGD) and the Jacobian-based saliency map attack (JSMA). The quantization-based transformation enhances PGD’s ability to apply global perturbations and improves JSMA’s focus on manipulating specific features. This is because vector quantization operates without relying on gradients, which can be a limitation in traditional gradient-based methods. This leads to improved attack success rates across various models and datasets. Integrating quantization into the adversarial attack pipeline not only enhances success rates but also preserves the imperceptibility of perturbations. Moreover, our approach enables dynamic control over perturbation levels, offering flexibility in tuning attack intensity to achieve desired outcomes. In summary, our method demonstrates the potential of using quantization to reinforce adversarial attacks, providing a novel direction for future research focused on improving the efficiency and robustness of attack strategies.