JavaScript-based attacks have become the major threat to web security. Because of complexities and obfuscation of JavaScript code, it is even more difficult to manually generate digital signatures, thus the traditional antivirus scanners can hardly detect malicious JavaScript effectively. In the paper, the characteristics of JavaScript code and extract parts of features to distinguish benign from malicious were analyzed, then we propose a method based on multilayer perceptron to detect malicious JavaScript code. This method can effectively detect the JavaScript-based attacks. Experimental results indicate that the average classification precision achieves 98.8% and the Average rate of false positives about 3%. The various performance indicators surpass the other machine learning-based methods.