The field of deep learning is moving towards improving adversarial robustness and incorporating geometric principles into model design. Researchers are exploring new methods to defend against adversarial attacks, such as gradient-feature alignment and radial compensation, which show promise in improving model robustness. Additionally, there is a growing interest in geometric deep learning, with studies on hyperbolic networks, Riemannian manifolds, and Kähler geometry. These advancements have the potential to lead to more robust and generalizable models. Noteworthy papers include: Transferable Hypergraph Attack via Injecting Nodes into Pivotal Hyperedges, which presents a novel framework for attacking hypergraph neural networks. Volatility in Certainty (VC): A Metric for Detecting Adversarial Perturbations, which introduces a new metric for detecting adversarial examples. Radial Compensation: Stable and Semantically Decoupled Generative Models on Riemannian Manifolds, which proposes a method for stable generative models on curved spaces.
Advances in Adversarial Robustness and Geometric Deep Learning
Sources
Transferable Hypergraph Attack via Injecting Nodes into Pivotal Hyperedges
Volatility in Certainty (VC): A Metric for Detecting Adversarial Perturbations During Inference in Neural Network Classifiers
Robust Bidirectional Associative Memory via Regularization Inspired by the Subspace Rotation Algorithm
CAO: Curvature-Adaptive Optimization via Periodic Low-Rank Hessian Sketching
Training Instabilities Induce Flatness Bias in Gradient Descent
Angular Gradient Sign Method: Uncovering Vulnerabilities in Hyperbolic Networks
DeepDefense: Layer-Wise Gradient-Feature Alignment for Building Robust Neural Networks
Certified but Fooled! Breaking Certified Defences with Ghost Certificates
Radial Compensation: Stable and Semantically Decoupled Generative Models on Riemannian Manifolds
Complex variational autoencoders admit K\"ahler structure