GENERATIVE ADVERSARIAL NETWORK WITH AUTOENCODER FOR CONTENT BASED IMAGE RETRIEVAL Subhra Samir Kundu Student, Amity Institute of Information Technology, Amity University Kolkata, Kolkata, 700135, India subhrasamirk@gmail.com Ambar Dutta Associate Professor, Amity Institute of Information Technology, Amity University Kolkata, Kolkata, 700135, India adutta@kol.amity.edu Abstract The internet generates a huge amount of information for a query, but not all of it is useful because it contains some misinformation and some manipulated data. Content-Based Image Retrieval (CBIR) is a state-of-the-art process that is employed by major IT companies all over the world. The research is nearly complete, and they are now being utilized to break the system rather than improving or classifying the right misinformation utilizing state-of-the-art adversarial networks. The major goal of this research is to classify a given misinformation and identify all of the images that were used to create it. A simple general adversarial network (GAN) is utilized in conjunction with an autoencoder to calculate the latent vector. Using the nearest neighbor computation metric, the latent vector is then used to obtain all of the closely matching images. Using the nearest neighbor computation metric, the latent vector is then used to obtain all of the closely matching images. The proposed study has demonstrated that it can retrieve images with much less distance than the current ones and those with a single component than using both in a collaboration. The proposal can lower the same in one-third of the cases already in use. Keywords: Content-Based Image Retrieval, General Adversarial Network, Autoencoder, Nearest Neighbor, Convolutional Neural Network. 1. Introduction Content-based Image Retrieval (CBIR) is a collection of approaches for retrieving semantically relevant photographs from a database using naturally inferred picture alternatives. Visual aspects are typically depicted at a low level in CBIR frameworks. They are essentially rigid numerical estimations that have no bearing on the subjectivity and haziness that characterize people's understandings and insights. As a result, there is a distinction to be made between low-level highlighting and unmistakable level semantics. One tends to observe a period of massive data processing where registering assets becomes the most significant bottleneck in dealing with such massive datasets. Because of the large dimensionality of data and the great spatiality of each perspective on it, feature selection is critical for improving clustering and classification outcomes. Due to the introduction of less expensive storage devices and, more importantly, the internet, extremely large collections of pictures are fast growing. Finding an image among a large number of images is a difficult undertaking. Physically naming images is one solution to this problem. Regardless, it is prohibitively expensive, time-consuming, and impractical for only a few applications. Furthermore, the naming interaction is dependent on the semantic accuracy of the image being depicted. As a result, many content-based image retrieval frameworks have been developed to extract low-level elements for displaying image content. Deep convolutional neural networks [Lai et al (2011)] have recently advanced the cutting edge in image categorization significantly, attracting a lot of attention in the computer vision field. The topic of image retrieval, for example, the task of identifying images that contain a comparable item or scene as in an inquiry picture, is related to the image classification problem. It has been suggested that features appearing in the upper layers of the CNN that figure out how to group photographs can serve as excellent descriptors for image recovery. The paper [Krizhevsky et al. (2012)] have demonstrated some subjective proof for the aforementioned. The biggest issue that has arisen as a result of the increase in pictorial data available on the internet is the problem of fraudulent images. These false images can be fully synthetic or man-made, or they can be created by e-ISSN : 0976-5166 p-ISSN : 2231-3850 Subhra Samir Kundu et al. / Indian Journal of Computer Science and Engineering (IJCSE) DOI : 10.21817/indjcse/2021/v12i6/211206146 Vol. 12 No. 6 Nov-Dec 2021 1780