Бучко, ОленаНгуєн, Сан Бинь2023-03-302023-03-302022Бучко О. А. Класифікація конфіденційних зображень із використанням нейронного хешу / Бучко О. А., Нгуєн С. Б. В. // Наукові записки НаУКМА. Комп'ютерні науки. - 2022. - Т. 5. - С. 68-71. - https://doi.org/10.18523/2617-3808.2022.5.68-712617-38082617-7323https://doi.org/10.18523/2617-3808.2022.5.68-71https://ekmair.ukma.edu.ua/handle/123456789/24865Humanity generates considerable information using its devices – smartphones, laptops, and tablets. Users upload images to different platforms, such as social networks, messengers, web services and other applications, which greatly endanger their personal information. User privacy has been exploited on the Internet for a long time. Interested parties lure potential customers into a trap of offers and services using such information as age, weight, nationality, religion and preferences. The sensitive information that may be contained in personal images is sometimes not recognized by their users as dangerous to share and, therefore, can easily be shared online by the owner without a second thought. This article inspects a neural hash algorithm for solving image classification tasks of confidential information and evaluates it via basic metrics. The main idea of the algorithm is to find similar images that will serve as an example for defining classes. The algorithm uses hash codes, ensuring users’ privacy. The evaluation of the algorithm is based on "The Visual Privacy (VISPR) Dataset". The main components of the algorithm are a neural network that generates vectors of extracted features for images and an indexed set of images (hash tables) that store knowledge about a particular domain. The critical aspect of the algorithm involves collisions of hash codes for similar images due to the similarity of their vectors of extracted features. The resulting hash codes can be identical or differ by a specific value of Hamming distance. Multiple hash tables with different hash functions are used to increase the recall or precision of the results. The effect of imperfect taxonomy was analyzed, which led to further filtration of abstract classes and increasing overall scores. Also, the article investigates the "pseudo-adaptivity" of the algorithm - the ability to classify new classes and add new cases to existing classes that were not included in the training stages. Such ability may be crucial for domains with many image instances or classes.Запропоновано підхід із застосуванням нейронного хешу для розв’язання задачі класифікації конфіденційних зображень. Головна ідея алгоритму полягає у пошуку схожих зображень – таких, що слугуватимуть взірцем для визначення класів. Алгоритм використовує хеш-коди, що дає змогу забезпечити приватність світлин користувачів. Розглянуто псевдоадаптивність мережі.ukнейронні мережіхешуваннянайближчі сусідинавчання на прикладахстаттяneural networkshashingnearest neighborinstance-based learningArticle