Phishing is a sort of online assault that endeavors to cheat delicate data of organization clients. Current phishing website page discovery strategies primarily utilize manual component assortment, and there are issues that include extraction is muddled and the conceivable connection between's highlights can't be stayed away from. To tackle the issues, another phishing site page location model is proposed, among which the primary segments are programmed taking in portrayals from multi-perspectives highlights through portrayal learning and extricating highlights by mixture profound learning organization. Initially, the model treats URL, HTML page substance, and document object model design of pages as character arrangements individually, also, utilizes portrayal learning innovation to consequently get familiar with the portrayal of the pages; at that point, sends different portrayals to a half breed profound learning network made out of a convolutional neural network also, a bidirectional long and transient memory network through various channels to separate neighborhood and worldwide highlights, and utilize the consideration component to fortify the influence of significant highlights; finally, the yield of various channels is combined to acknowledge classification expectation. Through four arrangements of trials to check the recognition impact of the model, the outcomes show that the general classification impact of the model is superior to the current exemplary phishing site page discovery strategies, the exactness arrives at 98.05%, and the bogus positive rate is just 0.25%. It is demonstrated that the techniques of extricating website page highlights from all angles through portrayal learning and cross breed profound learning organization can successfully improve the discovery impact of phishing site pages..