It is well known that once a Java application uses native C/C++ methods through the Java Native Interface (JNI), any security guarantees provided by Java might be invalidated by the native methods. So any vulnerability in this trusted native code can compromise the security of the Java program. Fuzzing test is an approach to software testing whereby the system being tested is bombarded with inputs generated by another program. When using fuzzer to test JNI programs, how to accurately reach the JNI functions and run through them to find the sensitive system APIs is the pre-condition of the test. In this paper, we present a heuristic fuzz generator method on JNI vulnerability detection based on the branch predication information of program. The result in the experiment shows our method can use less fuzzing times to reach more sensitive windows APIs in Java native code.