To cope with the problem of the low accuracy in detecting obfuscated malware, an algorithm to detect obfuscated malware based on boosting multi-level features is presented. After a disassembly analysis and static analysis for the obfuscated malware, the algorithm extracts features from three dimensions: opcode distribution, a function call graph, and a system call graph, which combines the statistic and semantic features to reflect the behavior characteristic of the malware, and then gives out the decision result based on weighted voting for a different feature analysis. It has been proven by experiment that the algorithms have a much higher accuracy on the testing dataset.