In bio-cryptography, biometric traits are replacing traditional passwords for secure exchange of cryptographic keys. The Fuzzy Vault (FV) scheme has been successfully employed to design bio-cryptographic systems as it can absorb a wide range of variation in biometric traits. Despite the intensity of research on FV based on physiological traits like fingerprints, iris, and face, there is no conclusive research on behavioral traits such as offline handwritten signature images, that have high inter-personal similarity and intra-personal variability. In this paper, a FV system based on the offline signature images is proposed. A two-step boosting feature selection (BFS) technique is proposed for selecting a compact and discriminant user-specific feature representation from a large number of feature extractions. The first step seeks dimensionality reduction through learning a population-based representation, that discriminates between different users in the population. The second step filters this representation to produce a compact user-based representation that discriminates the specific user from the population. This last representation is used to generate the FV locking/unlocking points. Representation variability is modeled by employing the BFS in a dissimilarity representation space, and it is considered for matching the unlocking and locking points during FV decoding. Proof of concept simulations involving 72,000 signature matchings (corresponding to both genuine and forged query signatures from the Brazilian Signature Database) have shown FV recognition accuracy of about 97% and system entropy of about 45-bits.