27 March 2024 in Expert Systems with Applications
17 October 2022 in Expert Systems with Applications
23 June 2025 in Expert Systems with Applications
20 May 2024 in IEEE Transactions on Big Data
Although face recognition starts to play an important role in our daily life, we need to pay attention that data-driven face recognition vision systems are vulnerable to adversarial attacks. However, current digital adversarial attacks and physical adversarial attacks both have drawbacks, with the former ones impractical and the latter one conspicuous, high-computational and low-executable. To address the issues, we propose a practical, executable, stealthy and low computational adversarial attack based on LED illumination modulation. To fool the systems, the proposed attack generates physically imperceptible luminance changes to human eyes through fast intensity modulation of scene LED illumination and uses the rolling shutter effect of CMOS image sensors in face recognition systems to implant luminance information perturbation to the captured face images. In summary, we present a denial-of-service (DoS) attack for face detection and an evasion attack for face verification. We also evaluate their effectiveness against well-known face detection models, Dlib, MTCNN and RetinaFace, and face verification models, Dlib, FaceNet, and ArcFace. The extensive physical experiments show that the success rates of DoS attacks against face detection models reach 97.67$\%$, 100$\%$, and 100$\%$, respectively, and the success rates of evasion attacks against all face verification models reach 100$\%$.
18 August 2025 in IEEE Internet of Things Journal
The reliability and robustness of lane detection play an instrumental role in the practical deployment of autonomous driving systems. Despite previous research indicating that adversarial examples can negatively affect lane detection models, leading to erroneous lane detection, most existing adversarial attacks are designed to attack lane detection models of a single deep learning-based task. To address this issue, we propose a toward cross-task physical adversarial attack via LED illumination modulation for lane detection systems (AdvLIM-LDs) by analyzing the common critical feature response mechanisms of lane detection models of different task types. The method first introduces a critical feature disruption (CFD) module to evaluate the impact of adversarial examples on critical features contributing to various tasks. Then, it utilizes the pulse width modulation of LEDs and the rolling shutter effect of CMOS image sensor to implant imperceptible adversarial perturbations during the image acquisition process. Finally, it enhances the cross-task transferability of adversarial examples by jointly optimizing a dual objective function composed of critical feature loss and model output accuracy loss. In digital domain simulations, adversarial examples generated by AdvLIM-LD caused an average detection accuracy degradation of 74.64% across lane detection models representing different task types, including the segmentation model SCNN, the anchor-based detector LaneATT, the curve-fitting model LSTR, and the keypoint detection model GANet. For cross-task attacks, the average accuracy degradation reached 55.37%. Physical-world experiments further validated the effectiveness of AdvLIM-LD, demonstrating its ability to degrade the source models’ average detection accuracy by over 80%.
24 December 2025 in Neural Networks