- Pose-Encoding Variational Autoencoder for Sign Language Generation Considering Diversity
Authors: Mohammed Illies Racar and Richard Bowden
Abstract: This paper addresses the problem of diversity-aware sign language generation, where, given an image (or sequence) of a signer, we generate another image with the same pose but different attributes (e.g. gender, skin color, etc.). For this purpose, we extend the variational inference paradigm to include information on conditioning on pose and attributes. This formulation improves the quality of the synthesized images. The generator framework is presented as a UNet architecture to ensure spatial preservation of the input pose and includes visual features from variational inference to maintain control over appearance and style. Each body part is generated with a separate decoder. This architecture allows the generator to deliver better results overall. Experiments on the SMILE II dataset show that the proposed model quantitatively outperforms state-of-the-art baselines in terms of diversity, pixel-wise image quality, and pose estimation. Quantitatively, it faithfully reproduces the non-manual features of signers.
2. Federated Learning for Fraud Detection Using Variational Autoencoders and Gaussian Mixture Models
Authors: Enrique Marmol Campos, Aurora González Vidal, José Luis Hernández Ramos, Antonio Scarmeta
Abstract: Federated Learning (FL) has become an attractive approach to collaboratively train machine learning (ML) models while preserving the privacy of data sources. However, most of the existing FL approaches are based on supervised techniques, which may require resource-intensive activities and human intervention to obtain labeled datasets. Moreover, in the scope of cyberattack detection, such techniques are unable to identify previously unknown threats. In this direction, this work proposes a novel unsupervised FL approach to identify potential misconduct in vehicular environments. It leverages the computing capabilities of public cloud services for model aggregation purposes and as a central repository of misconduct events, enabling vehicle-to-vehicle learning and collective defense strategies. Our solution integrates the use of Gaussian Mixture Models (GMM) and Variational Autoencoders (VAE) on the VeReMi dataset in a federated environment where each vehicle aims to train using only its own data. Additionally, we use Restricted Boltzmann Machines (RBM) for pre-training purposes and Fedplus as an aggregation function to increase the convergence of the model. Our approach typically provides superior performance (over 80 percent) compared to recent proposals based on supervised methods and artificial partitioning of his VeReMi dataset.
