random projections for biometrics security

I read the following in one of the research papers: Cancelable biometrics using random projections can be implemented by multiplying a random matrix with a feature vector. The result is a cancelable biometric template. y = M multiplied by x where y, M and x refer to the cancelable feature vector, Gaussian random projection matrix, and the feature vector extracted from the image, respectively. My question is: I think this is the same concept as compressive sensing, right? if yes, I know that compressive sensing is not a one-way transformation, I mean that the original image can be reconstructed from the compressed image. On the other hand, the cancelable technique must be a one-way transformation. Then, how can we say that the random projection explained by the above equation is a cancelable technique? Also, I need to know how can we know if the transformation one-way (the inverse operation is allowed) or not?

Thanks

[KALYAN ACHARJYA]

Please email questions to the authors of the mentioned paper, perhaps you will get a more clear answer.