Presenter: Tiziano Bianchi
Wednesday, June 22nd, 2022 17:00
Deep networks have shown remarkable performance in complex visual and classification tasks. However, their use in security sensitive applications raises serious concerns, as they can be targeted by adversaries. One of the most severe threats is represented by adversarial perturbations, a collection of methods that interfere with neural networks input data in order to produce undesired outputs.In this talk, I will present the Gaussian Class-Conditional Simplex (GCCS) loss, a novel approach for training deep networks with improved classification accuracy and adversarial robustness. The proposed method learns a mapping of the input classes onto Gaussian target distributions in the latent space, such that a hyperplane can be used as the optimal decision surface. Results show that GCCS provides improved robustness against adversarial perturbations, outperforming models trained with conventional adversarial training.
Biography: Tiziano Bianchi is an Associate Professor at Politecnico di Torino. From 2005 to 2012, he was with the Department of Electronics and Telecommunications, University of Florence, as a Research Assistant. He is a co-founder of ToothPic, a startup offering breakthrough technologies for camera identification. His research interests include multimedia security technologies, multimedia forensics, adversarial machine learning, signal processing in the encrypted domain, security aspects of compressed sensing.