Yufei Gao

Accurate medical image classification is a cornerstone of intelligent diagnostic systems; however, models must be robust to heterogeneous clinical data, sensitive to subtle pathological cues, and highly efficient for resource-constrained edge deployment. Existing lightweight networks often reduce computational cost by isolating channel groups, thereby weakening the representation of lesion features. Furthermore, they struggle to balance the suppression of shallow background noise with the preservation of deep semantic contours. To address these challenges, we propose TetraFuse, a synergistic four-dimensional dynamic fusion framework that integrates spatial, channel, statistical, and frequency cues. Specifically, a Cross-Channel Dynamic Aggregation module reconstructs global channel topology to prevent feature isolation with negligible computational overhead. To enhance representation fidelity across layers, a Local Variance-Guided Excitation mechanism suppresses shallow background artifacts by modeling statistical heterogeneity, while a High-Frequency Boundary Injector prevents over-smoothing by preserving deep contour details. Extensive experiments on three distinct medical datasets (COVID-19, ISIC 2018, and Kvasir) demonstrate that TetraFuse achieves state-of-the-art diagnostic performance. Notably, the TetraFuse-Tiny variant reduces floating-point operations (FLOPs) by 91.5% compared to the standard ResNet50 baseline. On the Kvasir endoscopy dataset, it achieves an accuracy of 0.913 and an AUC of 0.992 with a computational cost of only 0.345G FLOPs. These results confirm that TetraFuse successfully balances rigorous diagnostic precision with operational efficiency, making it highly suitable for intelligent clinical pattern recognition systems. Our codes will be made available to facilitate research and development.

As autonomous driving technology matures, evaluating and improving the robustness of autonomous vehicles is vital. While realistic and challenging long-tail traffic scenarios are crucial for training, their inherent rarity leaves even well-trained systems vulnerable. Although large-scale simulated scenario generation has been explored, it often suffers from a simulation-to-reality distribution gap, and existing methods lack a unified strategy to leverage real-world data to enhance robustness and safety. Complementary to prior work handling safety-critical scenarios at the policy level, we introduce a novel vulnerability-aware co-optimization framework designed to improve autonomous systems in the presence of collision-prone driving scenarios. Our framework establishes a closed-loop process between scenario generation and policy optimization, alternately exposing robustness vulnerabilities and refining policies over time. Specifically, we decompose real-world scenario generation into three tractable sub-tasks: pairwise collision checking, autonomous vehicle on-policy rollout estimation, and surrounding vehicle motion prior sampling. Simultaneously, we iteratively optimize driving policies within these scenarios while enforcing kinematic feasibility and cost constraints. Extensive experiments demonstrate the feasibility and generalizability of our proposed framework in exploring robustness vulnerabilities of autonomous driving systems.

With the proliferation of trajectory data generated by many Internet of Things (IoT) devices in the AI-driven transportation field, trajectory representation is crucial for extracting individual behavioral characteristics from intelligent IoT systems. Neural network-based trajectory representation learning methods excel at obtaining consistent representations of individual movements and mining spatio-temporal autocorrelation features of trajectories. However, existing methods achieve high accuracy in reliable experimental data, and their high performance is not always available in real-world wild scenarios, which is not available to reveal the performance bounds of trajectory representation learning models when confronted with real-world phenomena. Given this, we propose a universal trajectory adversarial attack framework that investigates the robustness vulnerabilities of trajectory representation learning models by generating adversarial trajectory examples. We reveal the impact of the attack surface in practical deployment and the adversarial perturbation budget on the trajectory adversarial attack performance. Guided by the new framework, we propose a universal black-box adversarial attack for trajectories, named Universal Trajectory Customized Iteration Attack (UTCIA). Specifically, we simulate trajectories through single-point perturbations to obtain a vulnerability-dependent set of victim points, selecting and perturbing only a small subset of points to degrade the entire model performance. Furthermore, we aggregate the discrete set of salient target trajectory points into a high-dimensional space for iterative direction estimation, and relax the constraint region to further compress noise. We generate adversarial trajectory examples for two trajectory representation downstream tasks with fundamentally different objectives using multiple large-scale real-world datasets. Extensive experiments demonstrate the feasibility and generalizability of our proposed framework in exploring the robustness vulnerabilities of trajectory representation learning models.