Semantic Smoothing via Novel View Synthesis for Robust SAR Image Classification

2026-05-14

AI Summary

This work addresses the vulnerability of deep neural networks to adversarial perturbations in synthetic aperture radar (SAR) automatic target recognition, where conventional random smoothing methods struggle to balance robustness and accuracy due to their use of isotropic noise that disrupts semantic image structure. To overcome this limitation, the authors propose a novel semantic smoothing approach that, for the first time, integrates semantics-preserving geometric transformations into the random smoothing framework. Specifically, they design a conditional novel-view synthesis model grounded in SAR imaging geometry to generate structured, semantically consistent transformations as substitutes for traditional noise, enabling robust aggregation of multi-view predictions. The method significantly enhances robustness against both general-purpose attacks (e.g., FGSM, PGD) and SAR-specific attacks (e.g., OTSA, SMGAA), while simultaneously improving classification accuracy on clean samples.

Abstract

Deep neural networks are vulnerable to adversarial perturbations, limiting deployment in safety-critical applications such as synthetic aperture radar (SAR) automatic target recognition (ATR). Randomized smoothing improves robustness by averaging predictions over noisy inputs, but isotropic noise often fails to preserve the semantic structure of SAR imagery. We propose semantic smoothing, a defense that replaces noised-based perturbations with structured randomized transformations generated by a novel view synthesis model. For SAR, we condition on acquisition geometry to synthesize multiple plausible radar views. Predictions across generated randomized views are aggregated to form a robust classifier. Experiments show that semantic smoothing improves robustness against standard attacks, such as FGSM and PGD, and SAR-specific attacks, such as OTSA and SMGAA, while also increasing clean classification accuracy. These results demonstrate that randomized smoothing via semantically preserving geometric transformations is a promising alternative to isotropic noise for adversarial defense in structured sensing domains.

Problem

• adversarial robustness
• SAR image classification
• randomized smoothing
• semantic structure
• automatic target recognition

Innovation

• semantic smoothing
• novel view synthesis
• SAR ATR
• adversarial robustness
• structured transformations