Filtered-ViT: A Robust Defense Against Multiple Adversarial Patch Attacks

The introduction of Filtered-ViT marks a significant advancement in the field of deep learning vision systems, particularly for applications in safety-critical domains such as healthcare. Traditional defenses often falter when faced with multiple adversarial patches, which can lead to misclassifications. Filtered-ViT addresses this vulnerability by integrating SMART Vector Median Filtering, allowing it to achieve 79.8% clean accuracy and 46.3% robust accuracy against four simultaneous 1% patches on ImageNet. This performance not only surpasses existing defenses but also demonstrates its capability to handle real-world challenges, as evidenced by its effectiveness in mitigating natural artifacts like occlusions and scanner noise in radiographic medical imagery. By being the first transformer to show unified robustness against both adversarial and naturally occurring disruptions, Filtered-ViT sets a new standard for reliability in critical applications, ensuring that diagnostic content r…