Phishing is a criminal deception of phishers intended to steal sensitive information such as username and password of financial account details in a pool of internet users and proliferate it through e-mails and social media. In this paper, we have proposed a machine learning based anti-phishing system based on lexical features, host properties with some evaluation techniques. The proposed system is trained using more than 10,000 phishing and licit URLs.  For accurate and reliable evaluation of our proposed system, we have used different types of SVM Kernel. Experimental results of our proposed system show more than 90% accuracy in detecting phishing site using SVM Kernel.