Abstract

Abstract
Case-based explanations are an intuitive method to gain insight into the decision-making process of deep learning models in clinical contexts. However, medical images cannot be shared as explanations due to privacy concerns. To address this problem, we propose a novel method for disentangling identity and medical characteristics of images and apply it to anonymize medical images. The disentanglement mechanism replaces some feature vectors in an image while ensuring that the remaining features are preserved, obtaining independent feature vectors that encode the images' identity and medical characteristics. We also propose a model to manufacture synthetic privacy-preserving identities to replace the original image's identity and achieve anonymization. The models are applied to medical and biometric datasets, demonstrating their capacity to generate realistic-looking anonymized images that preserve their original medical content. Additionally, the experiments show the network's inherent capacity to generate counterfactual images through the replacement of medical features.

Abstract

Abstract
With the ever-growing complexity of deep learning models for face recognition, it becomes hard to deploy these systems in real life. Researchers have two options: 1) use smaller models; 2) compress their current models. Since the usage of smaller models might lead to concerning biases, compression gains relevance. However, compressing might be also responsible for an increase in the bias of the final model. We investigate the overall performance, the performance on each ethnicity subgroup and the racial bias of a State-of-the-Art quantization approach when used with synthetic and real data. This analysis provides a few more details on potential benefits of performing quantization with synthetic data, for instance, the reduction of biases on the majority of test scenarios. We tested five distinct architectures and three different training datasets. The models were evaluated on a fourth dataset which was collected to infer and compare the performance of face recognition models on different ethnicity.

Abstract
Despite the high performance of current presentation attack detection (PAD) methods, the robustness to unseen attacks is still an under addressed challenge. This work approaches the problem by enforcing the learning of the bona fide presentations while making the model less dependent on the presentation attack instrument species (PAIS). The proposed model comprises an encoder, mapping from input features to latent representations, and two classifiers operating on these underlying representations: (i) the task-classifier, for predicting the class labels (as bona fide or attack); and (ii) the species-classifier, for predicting the PAIS. In the learning stage, the encoder is trained to help the task-classifier while trying to fool the species-classifier. Plus, an additional training objective enforcing the similarity of the latent distributions of different species is added leading to a 'PAI-species'-independent model. The experimental results demonstrated that the proposed regularisation strategies equipped the neural network with increased PAD robustness. The adversarial model obtained better loss and accuracy as well as improved error rates in the detection of attack and bona fide presentations. © 2019 Gesellschaft fuer Informatik.

Abstract
The application of machine learning algorithms to predict the mechanism of action (MoA) of drugs can be highly valuable and enable the discovery of new uses for known molecules. The developed methods are usually evaluated with small subsets of MoAs with large support, leading to deceptively good generalization. However, these datasets may not accurately represent a practical use, due to the limited number of target MoAs. Accurate predictions for these rare drugs are important for drug discovery and should be a point of focus. In this work, we explore different training strategies to improve the performance of a well established deep learning model for rare drug MoA prediction. We explored transfer learning by first learning a model for common MoAs, and then using it to initialize the learning of another model for rarer MoAs. We also investigated the use of a cascaded methodology, in which results from an initial model are used as additional inputs to the model for rare MoAs. Finally, we proposed and tested an extension of Mixup data augmentation for multilabel classification. The baseline model showed an AUC of 73.2% for common MoAs and 62.4% for rarer classes. From the investigated methods, Mixup alone failed to improve the performance of a baseline classifier. Nonetheless, the other proposed methods outperformed the baseline for rare classes. Transfer Learning was preferred in predicting classes with less than 10 training samples, while the cascaded classifiers (with Mixup) showed better predictions for MoAs with more than 10 samples. However, the performance for rarer MoAs still lags behind the performance for frequent MoAs and is not sufficient for the reliable prediction of rare MoAs.