POSTER: Universal Defensive Perturbations: A Cluster-Based Ensemble Defense
Year: 2026
Type: conference-paper
Abstract: Universal Adversarial Perturbations (UAPs) pose a major threat to deep neural networks by enabling input-agnostic attacks that cause widespread misclassification. Existing defenses like adversarial training are often computationally expensive and may degrade clean accuracy. We propose a min--max framework that learns a single Universal Defensive Perturbation (UDP) to counter multiple UAPs, and prove existence of such a perturbation. Since a single UDP generalizes poorly across diverse UAPs, we cluster UAPs by behavioral similarity and refine clusters using activation-space similarity, learning one defensive perturbation per cluster. To reduce training cost, we train each cluster's defense on its medoid representative, and introduce a Max-Confidence Medoid Defense (MMD) for inference-time selection against unseen UAPs. Evaluated on MNIST, CIFAR-10, CIFAR-100, and ImageNet, our cluster-based ensemble achieves strong robustness while preserving clean accuracy. Finally, we sketch a theoretical direction for future work based on anchor sets —a finite collection of perturbations that could provably cover the entire UAP space—and outline the open problems that stand between this theory and a practical algorithm.
Authors: Reeshav Chowdhury, Ayantika Chatterjee
Institution: Indian Institute of Technology Kharagpur
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225850985
