Kaixiang Yang

Decision Boundary Drift: A Security, Privacy, and Trust Risk of Continual Learning for Agentic AI in Edge Networks

Continual learning (CL) is a key paradigm that enables intelligent agents to operate autonomously in edge networks over the long term. However, continuous model updates can lead to catastrophic forgetting and representation instability in edge deployment scenarios, which may further induce Decision Boundary Drift (DBD). We propose a DBD-based adversarial attack framework that exploits class-level drift modeling and leverages the deformation of decision boundaries caused by incremental updates. We introduce multiple statistical metrics to quantify boundary drift, based on which class-level adversarial perturbations are constructed and further optimized in the input space to generate effective adversarial examples. Extensive experiments on multiple datasets and continual learning models demonstrate that the proposed method can significantly degrade model robustness, revealing non-negligible security risks in continuously evolving learning systems. Inspired by the security and trustworthiness requirements of edge intelligent agents, we systematically study and quantify DBD and its associated security risks in continual learning. Our findings reveal a practical yet underestimated attack surface and provide a foundation for future research on secure and robust continual learning systems.

ProtoPSR: Prototype-guided Pairwise Similarity Regularization for Robust Spatial Multi-Omics Integration

Spatially resolved multi-omics technologies measure multiple molecular layers at the same tissue locations, providing a more comprehensive view of tissue organization than any single modality. However, unsupervised spatial domain discovery remains challenging: noise across modalities and mismatched data distributions can distort neighborhood graphs, while pseudo-label self-training may drift over training, resulting in unstable optimization and fragmented domain maps. We propose ProtoPSR, an unsupervised framework for robust spatial multi-omics domain discovery. ProtoPSR stabilizes both local clustering structure and global pairwise relations. It integrates multiple modalities using a dual-graph backbone with stabilized dynamic graph refinement, and improves clustering via prototype-aware contrastive learning with confidence-guided updates. In addition, a pairwise similarity regularization aligns embedding similarities with pseudo same-cluster relations, improving global consistency. Extensive experiments on spatial multi-omics benchmarks show that ProtoPSR consistently outperforms representative baselines in clustering accuracy and pairwise agreement, and produces more spatially coherent domain maps. Moreover, ProtoPSR demonstrates strong robustness under modality-specific feature corruption.

FB-TDF1: A Fuzzy-Boundary, Timeliness- and Dispersion-Weighted Evaluation Metric for Anomaly Detection in Satellite Remote Sensing

Machine learning offers substantial potential for improving anomaly detection in satellite telemetry, a task central to spacecraft health monitoring. As modern satellites generate increasingly large volumes of multivariate telemetry, automated detection systems must evolve to reduce the monitoring burden on spacecraft operations engineers (SOEs) and mitigate operational risks. Although numerous time series anomaly detection (TSAD) methods have been proposed, reliably evaluating their performance under realistic telemetry conditions remains a persistent challenge. Recent transformer-based models have demonstrated strong capability in capturing long-range dependencies and multichannel interactions in telemetry and remote sensing data, thereby gaining increasing adoption in TSAD applications. However, these models exhibit characteristic behaviors—such as smooth attention-driven score transitions near event boundaries, sensitivity to weak precursor patterns leading to slight onset misalignment, and multi-head-induced isolated false alarms—that are not adequately handled by existing evaluation metrics. From an operational perspective, an effective metric should reward timely detection, tolerate the inherent ambiguity of expert-annotated anomaly boundaries, penalize dispersed false alarms that substantially increase SOE workload, and discourage the complete omission of anomalous events, since missing an entire spacecraft anomaly may lead to severe operational consequences even if the missed segment is short. To address these limitations, we propose fuzzy-boundary timeliness and dispersion-weighted F-score (FB-TDF1), a new evaluation metric that jointly accounts for i) detection timeliness, ii) boundary uncertainty in expert annotations, iii) the dispersion characteristics of false positives, and iv) event-level missed-anomaly risk through an event-aware recall correction. FB-TDF1 is specifically designed to reflect the behavioral patterns of modern transformer-based TSAD models and align with the practical evaluation needs of real satellite telemetry monitoring systems.

FedGRAIL: Trustworthy and Efficient Federated Class-Incremental Learning for Long-Tailed IoT Edge Intelligence