Classifying Google reCAPTCHA v2 - A study using transfer learning models and evaluating their robustness against adversarial perturbations

Sammanfattning: This thesis seeks to examine the suitability and robustness of transfer learning models in creating an efficient reCAPTCHA v2 classifier, and further evaluates their performance against various adversarial attacks. Three models - DenseNet201, EfficientNetV2, and InceptionV3 - were trained and assessed, highlighting the applicability of transfer learning techniques in the classification of reCAPTCHA v2 challenges. Despite variation in performance metrics, all models achieved satisfactory results, with DenseNet201 outperforming others in validation and test accuracy, and InceptionV3 demonstrating shortest training time. The paper additionally showed varying levels of robustness for several different adversarial attacks among the models, with EfficientNetV2 proving to be the most resilient. This variability, despite identical top layers, points to the underlying base architecture as a significant determinant of a model's robustness against adversarial attacks. Consequently, the study supports a comprehensive evaluation for model selection, considering not only the performance of metrics but also the underlying model’s architectural properties that may affect its robustness. Finally, this work indicates the potential of transfer learning models for image-based CAPTCHA challenge classification and stresses the need for further research focusing on enhancing the adversarial robustness of these models. The findings of this study contribute to the expanding body of research on transfer learning, showcasing its potential applications in the domain of image-based CAPTCHA systems.