|
|
|
| Robustness of Radar Intelligent Recognition Models Under Adversarial Samples Attacks |
| SHEN Tong1， CHEN Jingxian1， ZHONG Ping2 |
| 1. Shanghai Electro-Mechanical Engineering Institute， Shanghai 201109， China；
2. National Key Laboratory of Automatic Target Recognition （Changsha）， Changsha 410073， Hunan， China |
|
|
|
|
Abstract Addressing the problem of insufficient robustness of the intelligent recognition model of radar High Resolution Range Profile (HRRP) under adversarial sample attacks, a lightweight enhancement method was proposed in this study. Firstly, a comprehensive analysis was conducted using the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and a black-box migration attack to assess the vulnerability of the lightweight Convolutional Neural Network (CNN). Secondly, a cascaded defense strategy of “fast adversarial training + input denoising auto encoder + post-anomaly detection”was established. Finally, countermeasure experiments were carried out using three types of air targets and 6,000 sets of measured samples. The results show that this strategy can reduce the attack success rate to 9.2%, sacrificing only 2.1 percentage point of the cleaning accuracy, and increasing inference delay by less than 20%. It achieves a stable balance among model size, real-time performance, and robustness, providing a practical solution for the anti-interference design in radar-intelligent recognition systems.
|
|
Received: 11 November 2025
Published: 11 March 2026
|
|
|
|
|
|
|
|
|
2026, Vol. 9 
