Physical Multimodal Adversarial Attacks Against Object Detection Models in Industrial Monitoring Scenario

Deep Neural Networks (DNNs) have made significant advancements in computer vision, widely applied to various tasks. However, these models remain vulnerable to adversarial attacks. This study aims to reveal the adversarial threats faced by visible–infrared detection models in real-world scenarios and proposes a unified adversarial patch method, i.e., a single patch design effective across both visible and infrared modalities, based on a genetic algorithm. This method enables selective or balanced attacks on visible and infrared detectors, providing an in-depth analysis of the security of models in practical applications. Experimental results show that the method effectively reduces the detection model’s accuracy and demonstrates attack effects in simulated real-world scenarios. By optimizing the shape features of adversarial patches using a genetic algorithm and adaptively adjusting the attack strength across modalities via weight coefficients, the proposed method enhances the flexibility and robustness of cross-modal adversarial attacks. Additionally, the method employs the Expectation Transformation (EOT) strategy, showing strong robustness under different viewpoints. Extensive experiments validate the method’s effectiveness, with an attack success rate (ASR) exceeding 89%. This study provides a theoretical basis for improving the robustness and security of models and offers valuable insights for safety-critical applications such as intelligent surveillance.