<p>Surveillance systems play a vital role in ensuring public safety by detecting criminal activity, managing traffic, and more. Nowadays, deep learning (DL) and machine learning (ML) techniques are widely used in these systems to enhance their accuracy and efficiency. However, recent studies have shown that artificial intelligence (AI)-based systems, particularly those using ML and DL, are vulnerable to adversarial attacks, which can cause the model to make incorrect decisions. These attacks were originally designed for image models. In this study, we propose a new approach where adversarial attacks can be extended to real-time video surveillance systems. To demonstrate this, we applied our method to a real-time face mask detection system. The system is based on Multi-Task Cascaded Convolutional Networks (MTCNN) for face detection and MobileNet-v2 for face mask classification. Our pioneering framework shows how state-of-the-art adversarial attacks can be adapted for real-time surveillance systems. Experimental results show the impact of the adversarial attack, reducing the model’s performance from a precision (P) of 0.93, recall (R) of 0.93, F1 score (F) of 0.93, and accuracy (A) of 0.93 to just 0.22, 0.21, 0.22, and 0.22, respectively. This research highlights the vulnerabilities of critical video surveillance systems to adversarial threats, emphasizing the urgent need for strong defense mechanisms before these systems are deployed in real-world scenarios.</p>