This work proposes a novel privacy-preserving neural network feature representation to suppress the sensitive information of a learned space while maintaining the utility of the data. The new international regulation for personal data protection forces data controllers to guarantee privacy and avoid discriminative hazards while managing sensitive data of users. In our approach, privacy and discrimination are related to each other. Instead of existing approaches aimed directly at fairness improvement, the proposed feature representation enforces the privacy of selected attributes. This way fairness is not the objective, but the result of a privacy-preserving learning method. This approach guarantees that sensitive information cannot be exploited by any agent who process the output of the model, ensuring both privacy and equality of opportunity. Our method is based on an adversarial regularizer that introduces a sensitive information removal function in the learning objective. The method is evaluated on three different primary tasks (identity, attractiveness, and smiling) and three publicly available benchmarks. In addition, we present a new face annotation dataset with balanced distribution between genders and ethnic origins. The experiments demonstrate that it is possible to improve the privacy and equality of opportunity while retaining competitive performance independently of the task.

Face recognition (FR) systems have a growing effect on critical decision-making processes. Recent works have shown that FR solutions show strong performance differences based on the user’s demographics. However, to enable a trustworthy FR technology, it is essential to know the influence of an extended range of facial attributes on FR beyond demographics. Therefore, in this work, we analyze FR bias over a wide range of attributes. We investigate the influence of 47 attributes on the verification performance of two popular FR models. The experiments were performed on the publicly available MAAD-Face attribute database with over 120M high-quality attribute annotations. To prevent misleading statements about biased performances, we introduced control group-based validity values to decide if unbalanced test data causes the performance differences. The results demonstrate that also many nondemographic attributes strongly affect recognition performance, such as accessories, hairstyles and colors, face shapes, or facial anomalies. The observations of this work show the strong need for further advances in making the FR system more robust, explainable, and fair. Moreover, our findings might help to a better understanding of how FR networks work, enhance the robustness of these networks, and develop more generalized bias-mitigating FR solutions.

Several papers have recently appeared in the literature which propose pseudo-dynamic features for automatic static handwritten signature verification based on the use of gray level values from signature stroke pixels. Good results have been obtained using rotation invariant uniform local binary patterns LBP_8,1^riu2 plus LBP_16,2^riu2 and statistical measures from gray level co-occurrence matrices (GLCM) with MCYT and GPDS offline signature corpuses. In these studies the corpuses contain signatures written on a uniform white “nondistorting” background, however the gray level distribution of signature strokes changes when it is written on a complex background, such as a check or an invoice. The aim of this paper is to measure gray level features robustness when it is distorted by a complex background and also to propose more stable features. A set of different checks and invoices with varying background complexity is blended with the MCYT and GPDS signatures. The blending model is based on multiplication. The signature models are trained with genuine signatures on white background and tested with other genuine and forgeries mixed with different backgrounds. Results show that a basic version of local binary patterns (LBP) or local derivative and directional patterns are more robust than rotation invariant uniform LBP or GLCM features to the gray level distortion when using a support vector machine with histogram oriented kernels as a classifier.

In this paper we propose a new method for generating synthetic handwritten signature images for biometric applications. The procedures we introduce imitate the mechanism of motor equivalence which divides human handwriting into two steps: the working out of an effector independent action plan and its execution via the corresponding neuromuscular path. The action plan is represented as a trajectory on a spatial grid. This contains both the signature text and its flourish, if there is one. The neuromuscular path is simulated by applying a kinematic Kaiser filter to the trajectory plan. The length of the filter depends on the pen speed which is generated using a scalar version of the sigma lognormal model. An ink deposition model, applied pixel by pixel to the pen trajectory, provides realistic static signature images. The lexical and morphological properties of the synthesized signatures as well as the range of the synthesis parameters have been estimated from real databases of real signatures such as the MCYT Off-line and the GPDS960GraySignature corpuses. The performance experiments show that by tuning only four parameters it is possible to generate synthetic identities with different stability and forgers with different skills. Therefore it is possible to create datasets of synthetic signatures with a performance similar to databases of real signatures. Moreover, we can customize the created dataset to produce skilled forgeries or simple forgeries which are easier to detect, depending on what the researcher needs. Perceptual evaluation gives an average confusion of 44.06 percent between real and synthetic signatures which shows the realism of the synthetic ones. The utility of the synthesized signatures is demonstrated by studying the influence of the pen type and number of users on an automatic signature verifier.

This is the author’s version of a work that was accepted for publication in Pattern Recognition . Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Pattern Recognition , 48, 9 (2005) DOI: 10.1016/j.patcog.2015.03.019On-line signature verification still remains a challenging task within biometrics. Due to their behavioral nature (opposed to anatomic biometric traits), signatures present a notable variability even between successive realizations. This leads to higher error rates than other largely used modalities such as iris or fingerprints and is one of the main reasons for the relatively slow deployment of this technology. As a step towards the improvement of signature recognition accuracy, the present paper explores and evaluates a novel approach that takes advantage of the performance boost that can be reached through the fusion of on-line and off-line signatures. In order to exploit the complementarity of the two modalities, we propose a method for the generation of enhanced synthetic static samples from on-line data. Such synthetic off-line signatures are used on a new on-line signature recognition architecture based on the combination of both types of data: real on-line samples and artificial off-line signatures synthesized from the real data. The new on-line recognition approach is evaluated on a public benchmark containing both real versions (on-line and off-line) of the exact same signatures. Different findings and conclusions are drawn regarding the discriminative power of on-line and off-line signatures and of their potential combination both in the random and skilled impostors scenarios.M. D.-C. is supported by a PhD fellowship from the ULPGC and M.G.-B. is supported by a FPU fellowship from the Spanish MECD. This work has been partially supported by projects: MCINN TEC2012-38630- C04-02, Bio-Shield (TEC2012-34881) from Spanish MINECO, BEAT (FP7-SEC-284989) from EU, CECABANK and Cátedra UAM-Telefónic