Abstract-we proposed a novel approach for software trust analysis, in order to solve the problems in auto testing that, the test process is lack of guidance, further more, with high performance cost. The main idea of this approach is based on taint data tracing by taint tracing started from outside of software environment. By extracting the raw behaviors such as API invocation that may cause un-trust consequence, the scope of codes that being tested will be narrowed. These may-un-trust behaviors then form into a taint dependency behaviors model, and will be proved that whether these behaviors can be trusted in certain environment. The proving mechanism will be done by a so called noninterference information flow model. The behavior model will be evolved by refinement, and the context of each test iteration will be taken advantaged of for guidance. Besides, non-interference theories will also be applied in proving. We also briefly discussed about the implementation of this approach, and necessary theorems for this approach are also proved.