In the realm of biometric protection, the cancelable biometric template protection scheme holds a prominent position. By leveraging user-specific tokens and irreversible functions, it generates protected templates and conducts the authentication process within the template domain. However, existing schemes may be insecure when the tokens are stolen. In light of this, based on cosine Locality Sensitive Hashing (LSH) and the grouped inner product transformation, this paper meticulously designs and successfully implements a secure cancelable template authentication scheme for real-valued biometrics. Experiments are carried out on ear, face, and fingerprint datasets. The accuracy of this scheme is verified under both genuine-token and stolen-token scenarios, showing that the scheme can be applied to various biometric recognitions while maintaining high accuracy even when tokens are stolen. Moreover, through an analysis of the irreversibility of the scheme in the stolen-token scenario, it is confirmed that the scheme can effectively resist single-record attacks, multiple-record attacks, attacks based on inequalities and quadratic programming and known-sample attacks. The scheme is simple, efficient, and has high recognition accuracy and security, making it practical.