Dynamically Weighted Multi-View Semi-Supervised Learning for CAPTCHA.

With the development of Optical Character Recognition and artificial intelligence technologies, the security of Behavioral Completely Automated Public Turing test to tell Computers and Humans Apart (CAPTCHA) has become an increasingly difficult task. In order to prevent malicious attacks and maintain network security, most existing works on CAPTCHA are to construct a fine binary classifier model but are not yet capable of detecting new attack means during confrontation. This motivates us to propose a Dynamically Weighted Multi-View Semi-Supervised Learning, dubbed as DWMVSSL method, to relieve this problem. More specifically, our proposed method extracts hidden patterns from multiple perspectives and updates the view weighting dynamically which can constantly detect new attack means. In addition, due to existing some redundant feature in views, we design a Filter Artificial Bee Colony method, named as FABC for feature selection which can efficiently reduce the impact of high dimensional features. The experimental results show that, compared the existing representative baseline methods, our DWMVSSL method can effectively detecting new attacks on confrontation.