In biometric systems, a user’s template may be represented as a vector of dissimilarities from a cohort. This vector can be considered as a pseudonymous identifier in a template protection scheme. We show that for retinal graph templates, which are sparse and so lack regularity, the cohort-based representation achieves three key necessities for the template protection scheme proposed here. First, it has comparable accuracy to an unprotected scheme. Second, it is resistant to an effective inverse method that reconstructs biometric samples from comparison scores. Third, based on the “general framework for evaluating unlinkability” it has local and global linkability scores that are at least as low as well-known state-of-the-art template protection schemes. Our work demonstrates that when a retina is represented using a sparse template, it can be protected using the cohort-based representation.