Introduction
Deep Neural Networks (DNNs) have achieved remarkable success in various computer vision tasks but remain susceptible to adversarial examples. These are inputs specifically crafted with small, often imperceptible perturbations that lead models to incorrect predictions. While white-box attacks are highly effective when model parameters are known, black-box attacks—where the target model's architecture and weights are hidden—face significant challenges in transferability. This paper introduces the Restricted Region Based Iterative Gradient Method (RRI-GM) to enhance the transferability of adversarial examples by focusing perturbations on key object regions identified through object detection.
Methodology
The RRI-GM approach shifts the focus from global image perturbations to localized, object-aware modifications. The method consists of three primary phases:
Region Restriction
The first step involves identifying the most salient regions of the image. The authors utilize the YOLO V3 object detection algorithm to generate bounding boxes around primary objects. By restricting the attack to these regions, the method ensures that perturbations target the features most critical to the model's classification logic, rather than wasting noise on background pixels.
Initial Perturbation
Once regions are identified, an initial noise baseline is established. Among various techniques tested—including Gaussian noise and value-based additions—the "Average Perturbation" method proved most effective. This technique modifies pixels based on the average values of their neighbors, providing a robust starting point for gradient refinement.
Iterative Gradient Attack
The final phase employs iterative gradient-based methods using a surrogate white-box model (InceptionV3). The authors specifically leverage the Momentum Iterative Fast Gradient Sign Method (MI-FGSM) to refine the noise. Momentum helps stabilize the gradient direction, preventing the attack from becoming trapped in local maxima specific to the surrogate model, thereby improving transferability to unknown target models. The entire process is constrained by a maximum pixel change of 16 units to ensure the image remains visually "clean" to human observers.
Experimental Results
The effectiveness of RRI-GM was evaluated using 3,000 images from the ImageNet dataset (CAAD 2019 CTF). The attack was tested against five normally trained models and five defense-hardened models.
Key findings include:
- Normal Models: The accuracy of ResNetV2-152 plummeted from over 92% to approximately 5.86% when subjected to the MI-FGSM variant of RRI-GM.
- Defense Models: Even models specifically trained to resist adversarial attacks showed vulnerability. For example, AdvInceptionV3's accuracy dropped from 99.9% to roughly 28.2% after just three iterations of the attack.
- Localization Advantage: Results indicate that perturbing the object region is significantly more effective than whole-image perturbation, suggesting that background noise can sometimes dilute the impact of an adversarial attack.
Conclusion
The RRI-GM method demonstrates that object-aware localization combined with momentum-based iterative updates significantly enhances the transferability of adversarial examples in black-box scenarios. While the method depends on the accuracy of the underlying object detector and requires more computation than single-step attacks, it represents a substantial step forward in understanding and exploiting DNN vulnerabilities. Future work will focus on improving key region identification and optimizing the trade-off between attack stealthiness and success rates.
Summary updated