Solução de CAPTCHA baseado em imagem utilizando classificadores multiclasse e de classe única

Abstract

The use of CAPTCHA today is common for the protection of internet services to ensure their smooth and safe operation, however, it can be an inconvenience for some users who want to access these services correctly. One of these “users” are robots that, without any dishonest intention, are developed for the automation of processes in companies. One of the main types of CAPTCHA used is the image-based one, which consists of choosing the images of the requested object among different images, with Google’s reCAPTCHA being the most used. The objective of this work is to provide a model that is capable of solving these challenges without exploiting any technical vulnerability and that can be used by process automation tools. Three models were studied and trained: a multiclass, implemented in a convolutional neural network (CNN), and two of a single class known as One-Class SVM (OC-SVM) and kernel PCA (kPCA). The three models were trained with images availables on the internet and evaluated on a set of challenges formed by images obtained from reCAPTCHA itself. Four evaluation scenarios were elaborated, which differ in the way the images are chosen and in the criteria to determine if the challenge was solved. In the first two scenarios the CNN model performed better with 55% of the challenges solved. The kPCA and OC-SVM model had similar performance in the third scenario, solving 46 and 44% of the challenges, respectively. In the fourth and final scenario, kPCA solved 71% of the challenges, while OC-SVM only 65%. In these last scenarios the CNN model was not evaluated because it isn’t applicable to them.