PZPengfei Zhou

Pengfei Zhou

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Pengfei Zhou has a total of 46 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Resilient Path Tracking of Autonomous Driving under Few-shot Action Space Attacks

    by Yuting Wu, Xin Lou, Pengfei Zhou, Rui Tan, Zbigniew T. Kalbarczyk, Ravishankar K. Iyer

    Modern autonomous vehicles face growing cybersecurity risks, especially from action space attacks that directly target vehicle actuators. This article systematically evaluates the resilience of three representative Autonomous Driving (AD) architectures, including modular, end-to-end, and feature-fused agents, against few-shot action space attacks crafted via deep reinforcement learning under a black-box setting. The adversary perturbs the vehicle’s lateral control only during safety-critical moments, using either a camera or an inertial measurement unit. Our results reveal distinct vulnerabilities and behavioral patterns across AD architectures, which underscore the necessity for adaptive and robust defense strategies. However, existing adversarial training defense methods show limitations of overfitting and reliance on attack knowledge. To address these limitations, we propose a learning-based Path Correction System (PCS) that integrates traditional feedback control with an adversarially trained correction loop. The correction loop is selectively activated by a kinematic model-based attack detector to counteract abnormal control deviations. Evaluation experiments show that PCS reduces path-tracking deviation by 78% when the system is under attack.

  • Open Access

    A Two-Stage Selective Escalation Framework for Robust Suicide Risk Detection on Social Media: Machine Learning Study (Preprint)

    by Yukai Song, Pengfei Zhou, César Escobar-Viera, Candice Biernesser, Wei Huang, Jingtong Hu

    BACKGROUND Suicide rates have risen worldwide in recent years, underscoring the urgent need for proactive and scalable prevention strategies. Social media has emerged as a valuable source for early screening, as many at-risk individuals who avoid formal help due to stigma express distress online. However, current suicide risk detection systems face a critical limitation: while explicit suicidal expressions are relatively easy to identify, implicit suicidal ideation conveyed through metaphor, sarcasm, or subtle emotional cues remains difficult to detect and constitutes a major source of system failure in real-world deployment. OBJECTIVE This study aims to develop a robust and computationally efficient suicide risk detection framework that can reliably handle both explicit and implicit expressions while controlling inference cost and improving practical deployability for mental health screening. METHODS We propose a robust and efficient two-stage framework that explicitly accounts for heterogeneous signal strength through selective escalation. In Stage 1, a lightweight Bidirectional Encoder Representations from Transformers (BERT) classifier rapidly resolves high-confidence explicit cases. In Stage 2, only ambiguous inputs are escalated for further analysis through two complementary mechanisms: (i) a multi-perspective large language model (LLM) voting strategy designed to improve recall on implicit suicide risk, and (ii) a feature-based machine learning ensemble guided by psychologically grounded indicators extracted via prompt-engineered LLMs, enabling efficient inference and enhanced interpretability. RESULTS We evaluate the proposed framework on two complementary datasets representing distinct risk profiles: an explicit-dominant Reddit dataset and an implicit-only DeepSuiMind dataset. The framework achieves an F1 score of 98.0% on the Reddit benchmark and up to 99.94% recall on the implicit-only DeepSuiMind benchmark (positive-only; precision is not defined). It reduces the cross-domain performance gap to approximately 1.2-1.4% in recall and 1.7-2.2% in F1, and significantly lowers LLM invocation costs compared with single-model baselines. CONCLUSIONS These results demonstrate that system-level design with selective escalation offers a practical and effective pathway for suicide risk detection systems that are both robust to implicit expressions and efficient enough for real-world mental health screening applications. CLINICALTRIAL N/A

  • Open Access

    Invisible Adversarial Stripes on Traffic Sign: Threat and Defense for Autonomous Vehicles

    by Dongfang Guo, Yuting Wu, Pengfei Zhou, Xin Lou, Rui Tan

    Camera-based computer vision is essential to autonomous vehicle’s perception. This article presents an attack that uses light-emitting diodes and exploits the camera’s rolling shutter effect to create adversarial stripes in the captured images to mislead traffic sign recognition. The attack is stealthy because the stripes on the traffic sign are invisible to humans. For the attack to be threatening, the recognition results need to be stable over consecutive image frames. To achieve this, we design and implement GhostStripe , an attack system that controls the timing of the modulated light emission to adapt to camera operations and victim vehicle movements. Evaluated on real testbeds, GhostStripe can stably spoof the traffic sign recognition results for up to 94% of frames to a wrong class when the victim vehicle passes the road section. In reality, such attack effect may fool victim vehicles into life-threatening incidents. To counteract this threat, we propose GhostBuster , a software-based defense module to detect and mitigate the effects of GhostStripe. GhostBuster incorporates a perturbation detector and a sign restorer, effectively restoring the natural appearance of compromised traffic signs and significantly reducing the attack’s impact. We also discuss other countermeasures at the levels of camera sensor and autonomous driving system.

  • Open Access

    Large Language Model-guided Semantic Alignment for Human Activity Recognition

    by Hua Yan, Heng Tan, Yi Ding, Pengfei Zhou, Vinod Namboodiri, Yu Yang

    Human Activity Recognition (HAR) using Inertial Measurement Unit (IMU) sensors is critical for applications in healthcare, safety, and industrial production. However, variations in activity patterns, device types, and sensor placements create distribution gaps across datasets, reducing the performance of HAR models. To address this, we propose LanHAR, a novel system that leverages Large Language Models (LLMs) to generate semantic interpretations of sensor readings and activity labels for cross-dataset HAR. This approach not only mitigates cross-dataset heterogeneity but also enhances the recognition of new activities. LanHAR employs an iterative re-generation method to produce high-quality semantic interpretations with LLMs and a two-stage training framework that bridges the semantic interpretations of sensor readings and activity labels. This ultimately leads to a lightweight sensor encoder suitable for mobile deployment, enabling any sensor reading to be mapped into the semantic interpretation space. Experiments on five public datasets demonstrate that our approach significantly outperforms state-of-the-art methods in both cross-dataset HAR and new activity recognition. The source code is publicly available at https://github.com/DASHLab/LanHAR.


  • Open Access

    PICASSO 2025: 8th International Workshop on Physics Embedded AI Solutions in Mobile Computing

    by Julio Sahuquillo, Fan Dang, Miloš Stojmenović, Pengfei Zhou, Jingao Xu

  • Open Access

    Breaking the Resolution Barriers of mmWave Arrays via Null Steering for Sleep Monitoring in Multi-Person Scenarios

    by Duo Zhang, Xusheng Zhang, Zhehui Yin, Pengfei Zhou, Hongliu Yang, Pei Wang, Zhiyun Yao, Junzhe Wang, Fusang Zhang, Daqing Zhang

    Couples sharing a bed is a common practice in households, making sleep monitoring in multi-person scenarios a significant area of research. The high sensitivity of mmWave makes it a promising tool for monitoring vital signs during sleep. However, implementing mmWave-based solutions for such settings poses two major challenges. First, the range and angular resolution of commercial mmWave radar are still inadequate for handling such close-proximity scenarios where individuals share the same bed. Second, even with sufficient angular resolution, aligning the main lobe with one target usually faces strong interference from other targets located in the sidelobes. To address these challenges, we leverage antenna pattern nulls---directions with zero gain---to mitigate interference and overcome resolution limitations. Implementing null steering on practical mmWave radar, however, is challenging due to the human body being an extended target that spans a range of angles, as well as hardware imperfections that render traditional algorithms ineffective. In this paper, we present mmNull, the first null steering system designed for commercial mmWave radars. mmNull models the human body as an extended target rather than a single point, derives a closed-form solution to maximize target gain while suppressing interferences, and employs a heuristic algorithm to address hardware imperfections. Comprehensive experiments show that mmNull achieves a more than 5x improvement in accuracy compared to current state-of-the-art (SOTA) methods for monitoring respiration rate, heart rate, and heart rate variability (HRV) in multi-person scenarios.


  • Open Access

    Demo: Invisible Adversarial Stripes against Traffic Sign Recognition in Autonomous Driving

    by Dongfang Guo, Yuting Wu, Yimin Dai, Pengfei Zhou, Xin Lou, Rui Tan

    Camera-based computer vision is crucial for autonomous vehicle perception. We demonstrate GhostStripe [5], an attack system that uses light-emitting diodes and exploits the camera's rolling shutter effect to generate adversarial stripes that are invisible to humans while misleading traffic sign recognition. To maintain stable attack effectiveness, GhostStripe controls the timing of the modulated light emission, adapting to both the camera's framing operation and the movement of the victim vehicle. Evaluated on real testbeds, GhostStripe can stably spoof traffic sign recognition results for up to 97% of frames to a wrong class when the victim vehicle passes the road section.

  • Open Access

    Advancing PPG-Based Continuous Blood Pressure Monitoring from a Generative Perspective

    by Hui Ji, Pengfei Zhou

    Cuffless blood pressure (BP) monitoring is a critical task in the cardiovascular diseases (CVDs) domain, commonly based on Photoplethysmography (PPG) and Electrocardiogram (ECG) signals, providing foresight into cardiac health. While ECG often delivers better BP monitoring performance, the acquisition via straps and patches leads to a poor user experience. On the contrary, PPG enables continuous and convenient monitoring, but offers less informative references. A potential approach is to convert PPG signals into ECG signals, ensuring both high convenience and optimal accuracy. Converting PPG into ECG, however, involves a substantial reduction in inherent entropy, necessitating a thorough understanding of the process and specific techniques to guide the ECG generation. In this paper, we present a blood pressure monitoring framework that achieves ECG-level performance using solely the PPG signal. A diffusion model is introduced to conduct a selective ECG-targeted generative process with the condition of PPG. Based on our observation from the experimental investigation, a set of techniques is developed to significantly enhance the model's ability in generating high-quality ECG signals. Specifically, in the forward process, we employ an adaptive search module to adapt the QRS segment within the ECG waveform. In the reverse process, we propose the scale alignment and frequency alignment modules to better guide the generative process. Extensive experiments conducted on two public datasets and one self-collected dataset demonstrate the superior performance of our proposed framework, offering a groundbreaking perspective for PPG-based continuous blood pressure monitoring.