Friend Recommendation System using Image Encryption and Deep learning. Issue 1 (1st August 2022)

• Record Type: Journal Article
• Authors: Mohammed Yaseen, M.; Akilan, M.; Ranjith, A.G.; Soundariya, S.; Manisekaran, S.V.; Ramakrishnan, S.
• Is Part Of: Journal of physics. Volume 2325:Issue 1(2022)
• Publication Date: 2022-08-01
• Subjects: Friend recommendation system -- Deep learning -- Encryption; Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/2325/1/012044
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: The recommendation system functions similarly to a content filtering system, providing information depending on the preferences and behavior of the user. People use image sharing platforms such as Facebook, LinkedIn, and Instagram to share photographs of various elements to convey their sentiments with linked peers. As image sharing has been increasingly popular among individuals in recent years, it plays an important part in our daily lives. However, because of photo leaks and misuse, people are becoming increasingly concerned about their privacy. Encrypting photographs before publishing to social media is a better approach but decrypting the images and then searching for comparable images is computationally costly. The prior model advocated encrypting pictures using AES encryption and classifying them with deep autoencoders which is only a 2-step compression technique and takes quite more time for classifying. We provide a feasible privacy-protected friend recommendation system in proposed model by encrypting images using the Gaussian blur technique and a deep learning strategy for feature extraction and image classification for user recommendation. For training the dataset, the deep learning system utilizes the Keras model with CNN method, and a new model is generated from the features extracted from the blurred images. After that, the test dataset is compared to the model developed for similar image prediction. The accuracy measurements used to plot the graph are val acc and val loss. The accuracy result for the original image is 91.8%, while the accuracy result for the blurred image is 86.2% percent, demonstrating that our model is highly efficient and results in lower data loss.