Adversarial Robustness Using Transfer Learning Method: A Cost-Effective Approach

Remote sensing involves collecting data about objects or regions from a distance using satellite or aerial cameras. Deep learning models have achieved state-of-the-art performance in remote sensing by automatically extracting complex features, leading to more accurate predictions than traditional machine learning models. However, these deep learning models are highly vulnerable to adversarial attacks, which can significantly compromise their performance and reliability. This study uses adversarial patch attacks as the attack method, a physically realizable form of attack where a universal patch can be applied to any input to manipulate the model's predictions. An attacker can print this patch and place it on real-world objects, causing the model to misclassify images. Such attacks have proven highly effective in degrading the performance of deep learning models used in remote sensing. Adversarial training is a traditional defense approach against adversarial attacks, where models are trained using both genuine data and malicious inputs (adversarial examples). However, this process requires significant computational time. This study proposes a transfer learningbased strategy to reuse robust models, reducing the training time by approximately tenfold. We demonstrate that this transfer learning-based adversarial robustness technique achieves a 50% improvement in accuracy over the state-of-the-art models against adversarial patch attacks on benchmark UC-Merced and AID datasets.