ABSTRACT Hard disk drive (HDD) failures in large‐scale distributed storage systems can lead to severe data loss and service disruptions. While failure prediction using SMART data is a critical mitigation strategy, existing models often inadequately capture the gradual degradation of disk health, suffering from limitations in feature selection and temporal modeling. To overcome these challenges, this paper proposes GRUHP, a Gated Recurrent Unit‐based Health Prediction model integrated with an adaptive feature selection mechanism. GRUHP efficiently handles high‐dimensional Self‐Monitoring, Analysis, and Reporting Technology (SMART) data by dynamically identifying the most discriminative features, while its GRU architecture fully leverages temporal patterns for accurate health assessment. The model also incorporates dedicated modules for continuous health state evaluation and fault diagnosis. Extensive experiments on two public datasets demonstrate that GRUHP achieves an average precision of 93.4%, recall of 99.3%, and F 0.5 ‐score of 94.5%, with a false positive rate of only 0.7%. These results confirm that the proposed method, through its synergistic feature selection and temporal modeling, offers a robust and highly applicable solution for proactive failure prediction in real‐world storage environments.

ABSTRACT Serverless computing has emerged as a promising paradigm for cloud‐based stream processing applications characterized by fluctuating workloads and latency sensitivity. While existing Function‐as‐a‐Service (FaaS) implementations primarily focus on homogeneous CPU/memory resource scaling, they fail to address the challenges of heterogeneous resource management and coordinated elasticity in distributed stream processing. This study proposes HFaaS, a novel serverless framework that integrates dataflow programming with heterogeneous resource orchestration for stream processing applications. The key innovations include: (1) a dataflow‐oriented function composition model enabling dynamic scaling of individual processing stages through peer‐to‐point communication mechanisms, (2) a fine‐grained GPU resource allocation strategy achieving 15% + utilization improvement through device sharing and elastic scaling capabilities, and (3) a locality‐aware scheduling algorithm optimizing task placement based on data proximity and heterogeneous resource availability. Experimental results demonstrate that HFaaS effectively coordinates multi‐stage function scaling while maintaining sub‐second latency guarantees. The proposed resource allocation strategy improves GPU utilization by 15.2% compared to conventional static allocation approaches, with network overhead reduced by 31.6% through data‐local scheduling. This work bridges the gap between serverless architectures and modern stream processing requirements, providing a unified platform for building resource‐efficient, latency‐sensitive distributed applications in heterogeneous cloud environments.

ABSTRACT Amid the rapid expansion of digital infrastructure and the escalating sophistication of cyberattack strategies, network traffic anomaly detection has emerged as a critical cybersecurity mechanism for securing modern digital ecosystems. To overcome the shortcomings of traditional machine learning methods—specifically their limited accuracy in traffic pattern recognition—this paper proposes a novel C‐LSTM anomaly detection model enhanced by an attention mechanism. Building on advancements in deep learning architectures, the proposed model integrates CNNs and Bi‐LSTM networks to comprehensively capture spatial and temporal traffic features. The attention mechanism mitigates Bi‐LSTM's inherent vulnerability to vanishing gradients during long‐sequence data processing by adaptively reweighting feature significance, thereby optimizing detection performance. The model was rigorously validated using the NSL‐KDD and UNSW‐NB15 standard benchmark datasets and evaluated against contemporary state‐of‐the‐art detection methods. Experimental results demonstrate superior performance, with classification accuracies of 97.3% on NSL‐KDD and 95.8% on UNSW‐NB15, alongside a 12% reduction in false positives compared to baseline models. Notably, the attention mechanism achieved incremental accuracy improvements of 1.62% (NSL‐KDD) and 1.48% (UNSW‐NB15) compared to the baseline CNN‐LSTM model. These findings demonstrate the model's effectiveness in enhancing anomaly detection robustness, providing a practical framework for real‐world cybersecurity implementations.