On the limits of deep neural networks for CAPTCHA recognition under binary block encodings

Recent advances in deep learning have significantly weakened the security of traditional CAPTCHA systems, particularly those based on distorted text. In this work, we investigate the limits of neural CAPTCHA recognition by introducing a local binary block expansion scheme that replaces each pixel with a fixed-size binary pattern. By systematically controlling the Hamming density between foreground and background blocks, we analyze how representation-level changes affect the ability of modern sequence-based neural models to recognize characters. Extensive experiments using a convolutional recurrent neural network with CTC decoding demonstrate the existence of a sharp phase transition: while extreme-density encodings remain vulnerable, intermediate-density configurations cause recognition accuracy to collapse to chance level. Notably, this failure persists even when the attacker is fully informed and retrains the model on the modified CAPTCHA distribution. At the same time, human readability remains high for several resistant configurations. These results reveal a structural limitation of deep neural networks in learning from certain binary spatial representations and suggest a complementary representation-level transformation that can strengthen existing CAPTCHA schemes by adding an additional layer of robustness while preserving usability.