dc.description.abstract |
Abstract:
Our generation is quite obsessed with technology and we like to share our personal information such as photos and videos on the internet via different social networking websites i.e. Facebook, Snapchat, Instagram, etc. Therefore, it becomes easier for others to breach our privacy and harm us in a direct or indirect way. Now, computerized systems have advanced due to the improvements in Machine Learning (ML) algorithms and Artificial Intelligence (AI). These algorithms can extract sensitive information such as face attributes, text information, etc. from images or videos and can be used for privacy breaching. In this paper, we propose a novel privacy protection method by adding intelligent noise to the image while preserving image aesthetics and attributes. We determine multiple attributes for an image such as baldness, smiling, gender, etc. and we intelligently add noise to particular regions of the image that define a particular attribute using the visual explanation technique i.e. GradCam++, thereby preserving the other attributes. The addition of noise is based on the idea of Fast Gradient Sign Method (FGSM) that maximizes the gradients of the loss of an input image to create a new adversarial image. We integrate FGSM adversarial image and GradCam++ output to affect particular attributes only and hence keeping the image human imperceptible. The experiment results show that our attack outperforms the existing attacks including naive FGSM, Projected Gradient Descent (PGD), Momentum Iterative Method (MIM), Shadow Attack (SA), and Fast Minimum Norm (FMN) in terms of preserving attributes and image visual quality, when evaluated on CelebA dataset. |
en_US |