Pitfall of the Detection Rate Optimized Bit Allocation within template protection and a remedy

One of the requirements of a biometric template protection system is that the protected template ideally should not leak any information about the biometric sample or its derivatives. In the literature, several proposed template protection techniques are based on binary vectors. Hence, they require the extraction of a binary representation from the real-valued biometric sample. In this work we focus on the detection rate optimized bit allocation (DROBA) quantization scheme that extracts multiple bits per feature component while maximizing the overall detection rate. The allocation strategy has to be stored as auxiliary data for reuse in the verification phase and is considered as public. This implies that the auxiliary data should not leak any information about the extracted binary representation. Experiments in our work show that the original DROBA algorithm, as known in the literature, creates auxiliary data that leaks a significant amount of information. We show how an adversary is able to exploit this information and significantly increase its success rate on obtaining a false accept. Fortunately, the information leakage can be mitigated by restricting the allocation freedom of the DROBA algorithm. We propose a method based on population statistics and empirically illustrate its effectiveness. All the experiments are based on the MCYT fingerprint database using two different texture based feature extraction algorithms.