New NeurIPS 2025 Paper on Symmetry and Adversarial Robustness
A paper led by Dr. Longwei Wang, "Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness," appears in the Proceedings of the 39th Conference on Neural Information Processing Systems (NeurIPS 2025), one of the premier conferences in machine learning and artificial intelligence. The authors are Longwei Wang, Ifrat Ikhtear Uddin, KC Santosh, Chaowei Zhang, Xiao Qin, and Yang Zhou. Two of the authors, Dr. Wang and Dr. Chaowei Zhang, completed their Ph.D. degrees at Auburn University, co-advised by Dr. Xiao Qin and Dr. Ashish Gupta.
Deep neural networks can be fooled by adversarial examples: inputs altered by small, often imperceptible perturbations that cause a model to make wrong predictions. Adversarial training, the most widely used defense, is computationally expensive and can reduce a model's accuracy on clean data.
The paper takes an architectural approach instead. It embeds group-equivariant convolutions, specifically rotation- and scale-equivariant layers, into standard convolutional neural networks. These layers build symmetry priors into the model so that its behavior stays consistent under structured transformations of the input, which promotes smoother decision boundaries and greater resilience to attacks. The team proposes two symmetry-aware architectures: a parallel design that processes standard and equivariant features separately before fusing them, and a cascaded design that applies equivariant operations in sequence.
On the theoretical side, the paper shows that these models reduce hypothesis space complexity, regularize gradients, and yield tighter certified robustness bounds under the CLEVER (Cross Lipschitz Extreme Value for nEtwork Robustness) framework. In experiments on CIFAR-10, CIFAR-100, and CIFAR-10C, the symmetry-aware models consistently improve adversarial robustness and generalization against both FGSM and PGD attacks without requiring adversarial training. The results point to symmetry-enforcing architectures as efficient and principled alternatives to defenses based on data augmentation.
The work builds on Dr. Wang's doctoral research at Auburn University on understanding and improving the robustness of deep neural networks. Please read more about Dr. Wang's research here.
Click here to read the NeurIPS 2025 paper.