In recent years, public opinion attack has become more and more common. Attackers purposefully fabricate rumors and fake news, and these negative news widely spread across the Internet, which has a bad impact on individuals, companies or even the stock market. Most commonly, the malicious short-sellers attack targeted public companies by spreading rumors and disseminating distorted articles, resulting in stock price fluctuations or even sharp falls, to make profits therefrom. To protect personal reputation, the company's brand, and the stability of the stock market against public opinion attacks, it's so necessary to track fake news. However, rumors and fake news are mostly anonymous, and it lacks effective discriminating information to identify the authors, so it's so challenging to track these articles. Through the analysis of a large number of articles, it's found that different articles of the same author have a similar writing style, we can identify an anonymous article's author by writing style. We propose a unified method for authorship analysis based on deep learning, our method learns a mapping from texts to compact d-dimensional Euclidean space, this mapping embeds a text into the surface of a sphere with a radius of 1 and a center of origin. In Euclidean space, the distance indicates text similarity, texts of the same author have small distances, and texts of the distinct author have large distances. Inspired by mountain climbing, we propose a dynamic selection strategy to accelerate the model training, it can select efficient triplets before the beginning of each epoch. Finally, we construct a data set containing 130000 articles and 3600 authors for the experiment. The experimental results show that the dynamic selection strategy has a significant improvement, and our method has obvious advantages compared with other baselines, especially when the number of authors is large.