Timezone: »
(Non-)robustness of neural networks to small, adversarial pixel-wise perturbations, and as more recently shown, to even random spatial transformations (e.g., translations, rotations) entreats both theoretical and empirical understanding. Spatial robustness to random translations and rotations is commonly attained via equivariant models (e.g., StdCNNs, GCNNs) and training augmentation, whereas adversarial robustness is typically achieved by adversarial training. In this paper, we prove a quantitative trade-off between spatial and adversarial robustness in a simple statistical setting. We complement this empirically by showing that: (a) as the spatial robustness of equivariant models improves by training augmentation with progressively larger transformations, their adversarial robustness worsens progressively, and (b) as the state-of-the-art robust models are adversarially trained with progressively larger pixel-wise perturbations, their spatial robustness drops progressively. Towards achieving Pareto-optimality in this trade-off, we propose a method based on curriculum learning that trains gradually on more difficult perturbations (both spatial and adversarial) to improve spatial and adversarial robustness simultaneously.
Author Information
Sandesh Kamath (Indian Institute of Technology Hyderabad)
Amit Deshpande (Microsoft Research)
Subrahmanyam Kambhampati Venkata (Chennai Mathematical Institute)
Vineeth N Balasubramanian (Indian Institute of Technology, Hyderabad)
More from the Same Authors
-
2021 : ORDER: Open World Object Detection on Road Scenes »
Deepak Singh · Shyam Nandan Rai · Joseph K J · Rohit Saluja · Vineeth N Balasubramanian · Chetan Arora · Anbumani Subramanian · C.V. Jawahar -
2021 Poster: Adversarial Robustness without Adversarial Training: A Teacher-Guided Curriculum Learning Approach »
Anindya Sarkar · Anirban Sarkar · Sowrya Gali · Vineeth N Balasubramanian -
2020 : Contributed Talk 1: The Importance of Modeling Data Missingness in Algorithmic Fairness »
Naman Goel · Amit Deshpande -
2020 Poster: Meta-Consolidation for Continual Learning »
Joseph K J · Vineeth N Balasubramanian -
2016 Poster: Batched Gaussian Process Bandit Optimization via Determinantal Point Processes »
Tarun Kathuria · Amit Deshpande · Pushmeet Kohli