Many studies have demonstrated the vulnerability of Deep Neural Networks (DNNs) to adversarial attacks. While numerous research efforts have proposed high-performance adversarial attacks and defenses, there is a lack of research regarding the detection of defenses used by models. We have observed that different attacks have different effectiveness against the same defenses and that understanding defenses can help choose more effective attacks. In this paper, we propose a novel method for detecting defenses by generating defense detection examples encompassing gradient-based and decision-based algorithms. Defense detection examples are generated based on a specific defense. These examples do not exhibit any adversarial properties before undergoing the defense processing, but once subjected to the defense, adversarial properties become apparent. We can detect specific image-processing defenses, then select more effective attacks as well as enhance attacks against such defenses. We have conducted experiments on ImageNet, utilizing five models, and Cifar-10, utilizing six models, along with a total of nine defenses, and demonstrate the effectiveness of our method.

Excessive Data Exposure (EDE), where an API returns redundant data to the client beyond what is required for its functionality, has become a pervasive and severe security threat. However, automated detection techniques for such vulnerabilities remain underdeveloped, and existing methods, particularly black-box fuzzing, face significant bottlenecks in terms of accuracy and efficiency. To address these challenges, we propose Mockingbird, an automated detection tool based on a statically-assisted dynamic analysis approach. The tool leverages the JavaScript Proxy mechanism for efficient dynamic taint tracking to precisely identify the dangling data that is transmitted from an API response to the client but never consumed by any expected functionality, such as UI rendering or state management. Furthermore, to tackle the lack of a standardized benchmark in this domain, we have constructed and open-sourced EDEBench, the first persistent benchmark for EDE evaluation, comprising 8 popular open-source web projects built on diverse modern technology stacks. Experimental evaluation on EDEBench shows that, compared to the state-of-the-art, Mockingbird achieves an average F1-score improvement of 24.1% (Precision +15.8%, Recall +32.8%), enhances detection speed by nearly 20 times, and demonstrates broad applicability across all tested frameworks. These results provide a clear illustration of our tool’s accuracy, applicability, and efficiency. The source code is available at https://github.com/NeoSunJZ/Mockingbird-JS.

Many edge computing applications based on computer vision have harnessed the power of deep learning. As an emerging deep learning model for vision, Vision Transformer models have recently achieved record-breaking performance in various vision tasks. But many recent studies on the robustness of the Vision Transformer have shown that the Vision Transformer is still vulnerable to adversarial attacks and is easily affected by adversarial attacks, causing the model to misclassify the input. In this work, we ask an intriguing question: “Can Adversarial Perturbations against Vision Transformers be detected with model explanations?” Driven by this question, we observe that benign samples and adversarial examples have different attribution maps after applying the Grad-CAM interpretability method on the Vision Transformer model. We demonstrate that an adversarial example is a Feature Shift of the input data, which leads to an Attention Deviation of the visual model. We propose a framework for capturing the Attention Deviation of vision models to defend against adversarial attacks. Furthermore, experiments show that our model achieves expectative results.