Cloud-based tools have enabled users to expand their device storage while benefiting from integrated machine learning capabilities. However, this advancement has raised significant concerns regarding user privacy, particularly in the context of image data. Traditional privacy-preserving techniques such as encryption, watermarking, and steganography, though effective in earlier scenarios, are increasingly being undermined by the growing sophistication of deep learning (DL) models in image classification. These models can accurately classify complex and noisy data, rendering many conventional privacy-preserving methods insufficient. Moreover, traditional techniques often compromise usability by producing images that are difficult for humans to interpret. Many users, however, prefer to store and view their images in a human-understandable form, while preventing neural networks from accurately classifying their content. To address this challenge, we propose a privacy-preserving method that applies adversarial attacks targeted at specific objects within an image. By leveraging the seam carving technique, our method introduces perturbations that degrade classification model performance while preserving visual interpretability for human viewers. Our experiments show that the proposed method significantly reduces image classification accuracy without compromising image readability. Even at a moderate perturbation level of just 15%, our approach reduces classification accuracy by 75.29%, 75%, and 50.40% for three state-of-the-art models (ResNet50, VGG16, and EfficientNetB5) while maintaining human readability of the images. We further demonstrate that, for varying perturbations, the approach significantly degrades classification accuracy, offering a practical solution for preserving privacy in cloud-based image storage platforms.