SIFT-Graph: Benchmarking Multimodal Defense Against Image Adversarial Attacks With Robust Feature Graph

The introduction of SIFT-Graph marks a significant advancement in the field of artificial intelligence, particularly in enhancing the robustness of deep vision models against adversarial attacks. Traditional models are often vulnerable due to their reliance on pixel-level representations, which can be easily manipulated. SIFT-Graph addresses this issue by incorporating Scale-Invariant Feature Transform keypoints and a Graph Attention Network, allowing the model to capture resilient visual features that are less sensitive to perturbations. Preliminary results indicate that this framework effectively improves the robustness of models like Vision Transformers and Convolutional Neural Networks against gradient-based white box adversarial attacks, while only incurring a marginal drop in clean accuracy. This development is crucial as it not only fortifies AI systems against potential threats but also paves the way for more secure applications in various fields that depend on visual data.