A number of biometric characteristics exist for person identity verification. Each biometric has its strengths. However, they also suffer from disadvantages, for example, in the area of privacy protection. Security and privacy issues are becoming more important in the biometrics community. To enhance security and privacy in biometrics, cancellable biometrics have been introduced. In this paper, we propose cancellable biometrics for face recognition using an appearance based approach. Initially, an ICA coefficient vector is extracted from an input face image. Some components of this vector are replaced randomly from a Gaussian distribution which reflects the original mean and variance of the components. Then, the vector, with its components replaced, has its elements scrambled randomly. A new transformed face coefficient vector is generated by choosing the minimum or maximum component of multiple (two or more) differing cases of such transformed coefficient vectors. In our experiments, we compared the performance between the cases when ICA coefficient vectors are used for verification and when the transformed coefficient vectors are used for verification. We also examine the properties of changeability and reproducibility for the proposed method.