This paper proposes a semantic-based approach to malware behavioral signature extraction and detection. This approach extracts critical malware behaviors as well as dependencies among these behaviors, integrating instruction-level taint analysis and behavior-level semantics analysis. Then, it acquires anti-interference malware behavior signatures using anti-obfuscation engine to identify semantic irrelevance and semantically equivalence. Further, a prototype system based on this signature extraction and detection approach is developed and evaluated by multiple malware samples. Experimental results have demonstrated that the malware signatures extracted show good ability to anti obfuscation and the detection based on theses signatures could recognize malware variants effectively.