Deep neural networks (DNNs) have achieved remarkable success across vision tasks, yet they remain highly vulnerable to adversarial perturbations. While patch-based attacks have been explored as a localized and efficient adversarial attack strategy on DNNs, existing methods often require large perturbation areas or lack precise placement, making them either perceptually noticeable or less effective. To address these limitations, this paper proposes a gradient-guided adversarial patch attack that targets model-specific vulnerable regions. Using gradient-based sensitivity analysis, we are able to pinpoint the most important pixels in the input image that influence the decision of target DNNs. The backgrounds of these sensitive pixels are initially considered as patches. Subsequently, these patches are converted into imperceptible and highly lethal patches through an iterative alpha-controlled blending method. Extensive experiments on multiple benchmark datasets demonstrate that our patches achieve high attack success rates while covering only a small fraction of the original image. These findings underscore the vulnerability of DNNs to compact, well-placed adversarial patches and provide valuable insights for designing stronger defenses against such localized attacks. The source code of our work is available at the given link:  https://github.com/RishavKumarIIT/Gradient-Guided-Adversarial-Patch-Attack .