Semantic Smoothing Boosts SAR Image Classification Robustness
A novel defense technique known as semantic smoothing enhances the resilience of deep neural networks utilized in synthetic aperture radar (SAR) automatic target recognition (ATR). In contrast to randomized smoothing that relies on isotropic noise, semantic smoothing utilizes structured randomized transformations derived from an innovative view synthesis model based on acquisition geometry. This strategy generates various plausible radar views and combines predictions to create a strong classifier. Experimental results indicate enhanced resistance to conventional attacks (FGSM, PGD) as well as SAR-specific threats (OTSA, SMGAA), while also boosting clean classification accuracy. This method effectively tackles adversarial weaknesses in critical SAR applications.
Key facts
- Semantic smoothing replaces noise-based perturbations with structured randomized transformations from novel view synthesis.
- The model conditions on acquisition geometry to synthesize multiple radar views.
- Robustness improves against FGSM, PGD, OTSA, and SMGAA attacks.
- Clean classification accuracy also increases.
- Method targets safety-critical SAR automatic target recognition.
Entities
—