This letter proposes a threat discrimination methodology for distinguishing between sensor replay attacks and sensor bias faults, based on the specially designed watermark integrated with adaptive estimation. For each threat type, a watermark is designed based on the changes that the threat imposes on the system. Threat discrimination conditions are rigorously investigated to characterize quantitatively the class of attacks and faults that can be discriminated by the proposed scheme. A simulation is presented to illustrate the effectiveness of our approach.