Concetti Chiave
Proposing SAR-AE-SFP-Attack to generate real physics adversarial examples by altering scattering feature parameters.
Sintesi
1. Abstract:
- SAR target recognition models are vulnerable to adversarial examples.
- Current methods focus on 2D digital domain, lacking real physics considerations.
- SAR-AE-SFP-Attack generates real physics adversarial examples by altering scattering feature parameters.
2. Introduction:
- DNN-based SAR target recognition models face threats from adversarial examples.
- Existing methods include image and pseudo physics adversarial examples.
- Proposed SAR-AE-SFP-Attack enhances attack efficiency significantly.
3. Method:
- SAR-AE-SFP-Attack alters scattering feature parameters to generate adversarial examples.
- Utilizes RaySAR simulator for simulation and optimization.
- Innovative finite difference method for non-differentiable simulator.
4. Experiments and Analysis:
- Performance experiments show SAR-AE-SFP-Attack outperforms other methods.
- Hyperparameter experiments reveal impact of iteration numbers and target structure.
- Transfer experiments demonstrate transferability across viewpoints and models.
5. Conclusion:
- SAR-AE-SFP-Attack offers potential for physical world applications.
- Provides direction for future research in physical attacks.
Statistiche
"Experimental results show that our SAR-AE-SFP-Attack method demonstrates higher attack effectiveness compared to other general attack methods."
"The attack success rate achieved by iterating for 25 epochs was not significantly different and was even higher for certain classifiers."
"Adversarial examples generated for a specific classification model not only retained their attack efficacy when transferred to other models, but in some cases, even surpassed their performance on the original model."
Citazioni
"SAR-AE-SFP-Attack significantly improves attack efficiency on CNN-based models (over 30%) and Transformer-based models (over 13%)."
"The SAR-AE-SFP-Attack method adds perturbations directly to the target object, achieving a full-chain attack in the SAR imaging process."
"Extensive experiments show that our SAR-AE-SFP-Attack method demonstrates higher attack effectiveness compared to other general attack methods."