Biometric recognition systems are pattern-recognition systems designed to identify users based on a vector of features, which may pertain to physiological or behavioral characteristics. These features are typically stored in databases. Despite the encryption algorithms employed to secure these databases, there is always a chance, however small, that an attacker could decrypt the data and gain access to the full set of features. Subsequently, the attacker could compare the acquired biometric information with templates corresponding to user identities, potentially leading to unauthorized identification. This study focuses on leveraging convolutional neural networks (CNNs) to compress images in a manner that increases the difficulty of restoring them, thereby reducing the effectiveness of such attacks. We employ CNNs to generate deep representations (features) and utilize deep convolutional generative adversarial networks (DCGANs) to reconstruct the images from these features. Finally, the accuracy of the reconstructed images is analyzed using Bayesian beta regression and Random Forest to identify the CNN hyperparameters that have the greatest impact on reconstruction accuracy. Our findings indicate that for simpler datasets, such as black-and-white images of letters or digits (e.g., MNIST), the number of convolutional layers and the presence of pooling layers are the two most critical structural hyperparameters in securing image features. However, for more complex datasets, such as colored facial images (e.g., FaceScrub), activation functions and optimization algorithms become more significant in determining the security of biometric representations.

Expectation propagation (EP) is an approximate Bayesian inference (ABI) method which has seen widespread use across machine learning and statistics, owing to its accuracy and speed. However, it is often difficult to apply EP to models with complex likelihoods, where the EP updates do not have a tractable form and need to be calculated using methods such as multivariate numerical quadrature. These methods increase run time and reduce the appeal of EP as a fast approximate method. In this paper, we demonstrate that EP can still be made fast for certain models in this category. We focus on various types of linear regression, for which fast Bayesian inference is becoming increasingly important in the transition to big data. Fast EP updates are achieved through analytic integral reductions in certain moment computations. EP is compared to other ABI methods across simulations and benchmark datasets, and is shown to offer a good balance between accuracy and speed.