Robustifying Point Cloud Networks by Refocusing

The ability to cope with out-of-distribution (OOD) corruptions andadversarial attacks is crucial in real-world safety-demanding applications. Inthis study, we develop a general mechanism to increase neural networkrobustness based on focus analysis. Recent studies have revealed the phenomenon of \textit{Overfocusing}, whichleads to a performance drop. When the network is primarily influenced by smallinput regions, it becomes less robust and prone to misclassify under noise andcorruptions. However, quantifying overfocusing is still vague and lacks clear definitions.Here, we provide a mathematical definition of \textbf{focus},\textbf{overfocusing} and \textbf{underfocusing}. The notions are general, butin this study, we specifically investigate the case of 3D point clouds. We observe that corrupted sets result in a biased focus distribution comparedto the clean training set. We show that as focus distribution deviates from the one learned in thetraining phase - classification performance deteriorates. We thus propose a parameter-free \textbf{refocusing} algorithm that aims tounify all corruptions under the same distribution. We validate our findings on a 3D zero-shot classification task, achievingSOTA in robust 3D classification on ModelNet-C dataset, and in adversarialdefense against Shape-Invariant attack. Code is available in:https://github.com/yossilevii100/refocusing.