Zhide Chen
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
Zhide Chen has a total of 69 co-authors. Here are the top co-authors by number of shared publications:
- Xu Yang
Minjiang University
8 shared publications - Chen Feng
Department of Information Engineering, Fuzhou Polytechnic, Fuzhou 350108, China
7 shared publications - Kexin Zhu
Fujian Normal University
7 shared publications - Wencheng Yang
School of Mathematics, Physics and Computing, University of Southern Queensland, Toowoomba, QLD, Australia
5 shared publications Li Xu
Key Lab of Network Security and Cryptography, School of Mathematics and Computer Science Fujian Normal University Fujian Province Fuzhou China
4 shared publications- Xuechao Yang
School of Computing Technologies, RMIT University, Melbourne, VIC, Australia
4 shared publications
Latest Publications
- 8 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (14) , 30115-30132
Smart Lock Cybersecurity: A Comprehensive Review of Cross-Layer Threats and Defense Mechanisms
Smart locks have evolved from traditional mechanical devices into core security nodes in residential and enterprise IoT ecosystems, yet their cybersecurity remains a pressing global challenge. Existing research suffers from three fundamental limitations: isolated single-layer vulnerability analysis, over-reliance on idealized laboratory experiments, and a disproportionate focus on proposing defenses rather than evaluating their real-world effectiveness. To address these gaps, this paper presents a comprehensive cross-layer security review of smart locks, grounded in a systematic threat modeling framework Using the STRIDE methodology, we construct a three-dimensional threat model spanning hardware, communication, and application layers, and conduct quantitative risk assessment of 12 high-priority attack paths. We further analyze the amplification effect of cross-layer attack chains that have been largely overlooked in previous studies. Building on this threat model, we critically evaluate the state-of-the-art defensive technologies, including hardware/firmware protection, protocol hardening, and intelligent biometric authentication. We identify significant deployment barriers for promising techniques such as Physical Unclonable Functions (PUFs) and time-of-flight ranging, highlighting the fundamental tradeoff between security, power consumption, and cost. We systematically compare lock security standards, and reveal a critical disconnect between academic research and industry practice. To bridge this gap, we propose a full-lifecycle security due diligence framework tailored for consumers, manufacturers, and regulators. This study provides both theoretical foundations and actionable guidelines for building end-to-end secure smart lock systems. Future research should prioritize low-power cross-layer attack detection, generalizable biometric anti-spoofing, and supply chain security for resource-constrained IoT devices.
- 8 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (14) , 30115-30132
Smart Lock Cybersecurity: A Comprehensive Review of Cross-Layer Threats and Defense Mechanisms
Smart locks have evolved from traditional mechanical devices into core security nodes in residential and enterprise IoT ecosystems, yet their cybersecurity remains a pressing global challenge. Existing research suffers from three fundamental limitations: isolated single-layer vulnerability analysis, over-reliance on idealized laboratory experiments, and a disproportionate focus on proposing defenses rather than evaluating their real-world effectiveness. To address these gaps, this paper presents a comprehensive cross-layer security review of smart locks, grounded in a systematic threat modeling framework Using the STRIDE methodology, we construct a three-dimensional threat model spanning hardware, communication, and application layers, and conduct quantitative risk assessment of 12 high-priority attack paths. We further analyze the amplification effect of cross-layer attack chains that have been largely overlooked in previous studies. Building on this threat model, we critically evaluate the state-of-the-art defensive technologies, including hardware/firmware protection, protocol hardening, and intelligent biometric authentication. We identify significant deployment barriers for promising techniques such as Physical Unclonable Functions (PUFs) and time-of-flight ranging, highlighting the fundamental tradeoff between security, power consumption, and cost. We systematically compare lock security standards, and reveal a critical disconnect between academic research and industry practice. To bridge this gap, we propose a full-lifecycle security due diligence framework tailored for consumers, manufacturers, and regulators. This study provides both theoretical foundations and actionable guidelines for building end-to-end secure smart lock systems. Future research should prioritize low-power cross-layer attack detection, generalizable biometric anti-spoofing, and supply chain security for resource-constrained IoT devices.
- Open Access
- 23 March 2026
- Published by MDPI AG in Electronics
- Vol. 15 (6) , 1331
Remote Sensing Identification of Benggang Using a Two-Stream Network with Multimodal Feature Enhancement and Sparse Attention
Benggang (Benggang), a typical landform characterized by severe erosion and a geohazard in the red-soil hilly regions of southern China, is characterized by a fragmented texture, irregular boundaries, and high similarity to background objects such as bare soil and roads, which poses a dual challenge of “multiscale variability + strong noise” for automated identification at regional scales. To address insufficient information from a single modality and the limited representation of cross-scale features, this study proposes a dual-stream feature-fusion network (DF-Net) for multisource data consisting of a digital orthophoto map (DOM) and a digital elevation model (DEM). The method adopts ResNeSt50d as the backbone of the two branches: on the DOM side, a Canny-edge channel is stacked to enhance high-frequency boundary information; on the DEM side, derived terrain factors, including slope, aspect, curvature, and hillshade, are introduced to provide morphological constraints. In the cross-modal fusion stage, a multiscale sparse attention fusion module is designed, which acquires contextual information via multiwindow average pooling and suppresses noise interference through top-K sparsification. In the decision stage, a multibranch ensemble is employed to improve classification stability. Taking Anxi County, Fujian Province, as the study area, a coregistered dataset of GF-2 (1 m) DOM and ALOS (12.5 m) DEMs is constructed, and a zonal partitioning strategy is adopted to evaluate the model’s generalization ability. The experimental results show that DF-Net achieves 97.44% accuracy, 85.71% recall, and an 82.98% F1 score in the independent test zone, outperforming multiple mainstream CNN/transformer classification models. This study indicates that the strategy of “multimodal feature enhancement + sparse attention fusion” tailored to Benggang erosional landforms can significantly improve recognition performance under complex backgrounds, providing technical support for rapid Benggang surveys and governance-effectiveness assessments.
- 23 March 2026
- Published by MDPI AG in Electronics
- Vol. 15 (6) , 1331
Open AccessRemote Sensing Identification of Benggang Using a Two-Stream Network with Multimodal Feature Enhancement and Sparse Attention
Benggang (Benggang), a typical landform characterized by severe erosion and a geohazard in the red-soil hilly regions of southern China, is characterized by a fragmented texture, irregular boundaries, and high similarity to background objects such as bare soil and roads, which poses a dual challenge of “multiscale variability + strong noise” for automated identification at regional scales. To address insufficient information from a single modality and the limited representation of cross-scale features, this study proposes a dual-stream feature-fusion network (DF-Net) for multisource data consisting of a digital orthophoto map (DOM) and a digital elevation model (DEM). The method adopts ResNeSt50d as the backbone of the two branches: on the DOM side, a Canny-edge channel is stacked to enhance high-frequency boundary information; on the DEM side, derived terrain factors, including slope, aspect, curvature, and hillshade, are introduced to provide morphological constraints. In the cross-modal fusion stage, a multiscale sparse attention fusion module is designed, which acquires contextual information via multiwindow average pooling and suppresses noise interference through top-K sparsification. In the decision stage, a multibranch ensemble is employed to improve classification stability. Taking Anxi County, Fujian Province, as the study area, a coregistered dataset of GF-2 (1 m) DOM and ALOS (12.5 m) DEMs is constructed, and a zonal partitioning strategy is adopted to evaluate the model’s generalization ability. The experimental results show that DF-Net achieves 97.44% accuracy, 85.71% recall, and an 82.98% F1 score in the independent test zone, outperforming multiple mainstream CNN/transformer classification models. This study indicates that the strategy of “multimodal feature enhancement + sparse attention fusion” tailored to Benggang erosional landforms can significantly improve recognition performance under complex backgrounds, providing technical support for rapid Benggang surveys and governance-effectiveness assessments.
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP (00983063) , 1
An LLM-Enhanced Weakly Supervised Framework for Securing Video Anomaly Detection in Consumer Devices
With the evolution of artificial intelligence (AI) and the Internet of Things (IoT) in urban environments, smart surveillance has become a vital security tool; it prevents minor irregularities from spiraling into major risks for both citizens and city systems. Video anomaly detection (VAD) confronts dual challenges: the complexity of anomalous events and the scarcity of labeled data, further compounded by environmental noise in real-world deployments. To address these issues, we propose a weakly supervised VAD framework that leverages Large Language Models (LLMs) for enhanced semantic understanding. The proposed model integrates Sequential Context-Aware Temporal (SCAT) modeling to capture spatiotemporal patterns while suppressing noise through refined attention mechanisms, and Multi-Perspective Semantic Label Generation (MPSLG) to produce semantically rich labels via LLM-based prompt engineering. SCAT strengthens temporal feature learning and noise resilience, while MPSLG provides high-quality semantic supervision, jointly enabling accurate and interpretable anomaly detection. Extensive experiments demonstrate substantial performance gains, achieving 83.82% average precision on XD-Violence for coarse-grained detection and 30.79% average mean average precision for fine-grained localization, outperforming state-of-the-art methods and validating effectiveness in both urban surveillance and consumer applications.
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP (00983063) , 1
An LLM-Enhanced Weakly Supervised Framework for Securing Video Anomaly Detection in Consumer Devices
With the evolution of artificial intelligence (AI) and the Internet of Things (IoT) in urban environments, smart surveillance has become a vital security tool; it prevents minor irregularities from spiraling into major risks for both citizens and city systems. Video anomaly detection (VAD) confronts dual challenges: the complexity of anomalous events and the scarcity of labeled data, further compounded by environmental noise in real-world deployments. To address these issues, we propose a weakly supervised VAD framework that leverages Large Language Models (LLMs) for enhanced semantic understanding. The proposed model integrates Sequential Context-Aware Temporal (SCAT) modeling to capture spatiotemporal patterns while suppressing noise through refined attention mechanisms, and Multi-Perspective Semantic Label Generation (MPSLG) to produce semantically rich labels via LLM-based prompt engineering. SCAT strengthens temporal feature learning and noise resilience, while MPSLG provides high-quality semantic supervision, jointly enabling accurate and interpretable anomaly detection. Extensive experiments demonstrate substantial performance gains, achieving 83.82% average precision on XD-Violence for coarse-grained detection and 30.79% average mean average precision for fine-grained localization, outperforming state-of-the-art methods and validating effectiveness in both urban surveillance and consumer applications.
- Open Access
- 30 December 2025
- Published by MDPI AG in Electronics
- Vol. 15 (1) , 179
BERTSC: A Multi-Modal Fusion Framework for Stablecoin Phishing Detection Based on Graph Convolutional Networks and Soft Prompt Encoding
As stablecoins become increasingly prevalent in financial crimes, their usage for illicit activities has reached a scale of USD 51.3 billion. Detecting phishing activities within stablecoin transactions has emerged as a critical challenge in blockchain security. Currently, existing detection methods predominantly target mainstream cryptocurrencies like Ethereum and lack specialized models tailored to the unique transaction patterns of stablecoin networks. This paper introduces a deep learning framework, BERTSC, based on multi-modal fusion. The model integrates three core modules graph convolutional networks (GCNs), BERT semantic encoders, and soft prompt encoders to identify malicious accounts. The GCN constructs directed multi-graph representations of account interactions, incorporating multi-dimensional edge features; the BERT encoder transforms discrete transaction attributes into semantically rich continuous vector representations; the soft prompt encoder maps account interaction features into learnable prompt vectors. An innovative three-way gated dynamic fusion mechanism optimally combines the information from these sources. The fused features are then classified to predict phishing account labels, facilitating the detection of phishing scams in stablecoin transaction datasets. Experimental results on large-scale stablecoin datasets demonstrate that BERTSC outperforms baseline models, achieving improvements of 4.96%, 3.60%, and 4.23% in Precision, Recall, and F1-score, respectively. Ablation studies validate the effectiveness of each module and confirm the necessity and superiority of the three-way gating fusion mechanism. This research offers a novel technical approach for phishing detection within blockchain stablecoin ecosystems.
- 30 December 2025
- Published by MDPI AG in Electronics
- Vol. 15 (1) , 179
Open AccessBERTSC: A Multi-Modal Fusion Framework for Stablecoin Phishing Detection Based on Graph Convolutional Networks and Soft Prompt Encoding
As stablecoins become increasingly prevalent in financial crimes, their usage for illicit activities has reached a scale of USD 51.3 billion. Detecting phishing activities within stablecoin transactions has emerged as a critical challenge in blockchain security. Currently, existing detection methods predominantly target mainstream cryptocurrencies like Ethereum and lack specialized models tailored to the unique transaction patterns of stablecoin networks. This paper introduces a deep learning framework, BERTSC, based on multi-modal fusion. The model integrates three core modules graph convolutional networks (GCNs), BERT semantic encoders, and soft prompt encoders to identify malicious accounts. The GCN constructs directed multi-graph representations of account interactions, incorporating multi-dimensional edge features; the BERT encoder transforms discrete transaction attributes into semantically rich continuous vector representations; the soft prompt encoder maps account interaction features into learnable prompt vectors. An innovative three-way gated dynamic fusion mechanism optimally combines the information from these sources. The fused features are then classified to predict phishing account labels, facilitating the detection of phishing scams in stablecoin transaction datasets. Experimental results on large-scale stablecoin datasets demonstrate that BERTSC outperforms baseline models, achieving improvements of 4.96%, 3.60%, and 4.23% in Precision, Recall, and F1-score, respectively. Ablation studies validate the effectiveness of each module and confirm the necessity and superiority of the three-way gating fusion mechanism. This research offers a novel technical approach for phishing detection within blockchain stablecoin ecosystems.
- 18 December 2025
- Published by Springer Nature in Empirical Software Engineering
- Vol. 31 (3) , 51
eBPF-Guard: a detection method for container escape via multi-level monitoring and enhanced analysis model
- 18 December 2025
- Published by Springer Nature in Empirical Software Engineering
- Vol. 31 (3) , 51
eBPF-Guard: a detection method for container escape via multi-level monitoring and enhanced analysis model
- 11 December 2025
- Published by Elsevier in Journal of Information Security and Applications
- Vol. 97
Anomaly detection for blockchain nodes based on eBPF and fine-tuning large language model
- 11 December 2025
- Published by Elsevier in Journal of Information Security and Applications
- Vol. 97
Anomaly detection for blockchain nodes based on eBPF and fine-tuning large language model
- Open Access
- 26 November 2025
- Published by MDPI AG in Preprints.org
BERTSC: A Multi-Modal Fusion Framework for Stablecoin Phishing Detection Based on Graph Convolutional Networks and Soft Prompt Encoding
As stablecoins become increasingly prevalent in financial crimes, their usage for illicit activities has reached a scale of \$51.3 billion. Detecting phishing activities within stablecoin transactions has emerged as a critical challenge in blockchain security. Currently, existing detection methods predominantly target mainstream cryptocurrencies like Ethereum and lack specialized models tailored to the unique transaction patterns of stablecoin networks. This paper introduces a deep learning framework, BERTSC, based on multimodal fusion. The model integrates three core modules graph convolutional networks (GCNs), BERT semantic encoders, and soft prompt encoders to identify malicious accounts. The GCN constructs directed multi-graph representations of account interactions, incorporating multi-dimensional edge features; the BERT encoder transforms discrete transaction attributes into semantically rich continuous vector representations; the soft prompt encoder maps account interaction features into learnable prompt vectors. An innovative three-way gated dynamic fusion mechanism optimally combines the information from these sources. The fused features are then classified to predict phishing account labels, facilitating the detection of phishing scams in stablecoin transaction datasets. Experimental results on large-scale stablecoin datasets demonstrate that BERTSC outperforms baseline models, achieving improvements of 4.96\%, 3.60\%, and 4.23\% in Precision, Recall, and F1-score, respectively. Ablation studies validate the effectiveness of each module and confirm the necessity and superiority of the three-way gating fusion mechanism. This research offers a novel technical approach for phishing detection within blockchain stablecoin ecosystems.
- 26 November 2025
- Published by MDPI AG in Preprints.org
Open AccessBERTSC: A Multi-Modal Fusion Framework for Stablecoin Phishing Detection Based on Graph Convolutional Networks and Soft Prompt Encoding
As stablecoins become increasingly prevalent in financial crimes, their usage for illicit activities has reached a scale of \$51.3 billion. Detecting phishing activities within stablecoin transactions has emerged as a critical challenge in blockchain security. Currently, existing detection methods predominantly target mainstream cryptocurrencies like Ethereum and lack specialized models tailored to the unique transaction patterns of stablecoin networks. This paper introduces a deep learning framework, BERTSC, based on multimodal fusion. The model integrates three core modules graph convolutional networks (GCNs), BERT semantic encoders, and soft prompt encoders to identify malicious accounts. The GCN constructs directed multi-graph representations of account interactions, incorporating multi-dimensional edge features; the BERT encoder transforms discrete transaction attributes into semantically rich continuous vector representations; the soft prompt encoder maps account interaction features into learnable prompt vectors. An innovative three-way gated dynamic fusion mechanism optimally combines the information from these sources. The fused features are then classified to predict phishing account labels, facilitating the detection of phishing scams in stablecoin transaction datasets. Experimental results on large-scale stablecoin datasets demonstrate that BERTSC outperforms baseline models, achieving improvements of 4.96\%, 3.60\%, and 4.23\% in Precision, Recall, and F1-score, respectively. Ablation studies validate the effectiveness of each module and confirm the necessity and superiority of the three-way gating fusion mechanism. This research offers a novel technical approach for phishing detection within blockchain stablecoin ecosystems.
- Open Access
An Interpretable Soft-Sensor Framework for Dissertation Peer Review Using BERT
Graduate education has entered the era of big data, and systematic analysis of dissertation evaluations has become crucial for quality monitoring. However, the complexity and subjectivity inherent in peer-review texts pose significant challenges for automated analysis. While natural language processing (NLP) offers potential solutions, most existing methods fail to adequately capture nuanced disciplinary criteria or provide interpretable inferences for educators. Inspired by soft-sensor, this study employs a BERT-based model enhanced with additional attention mechanisms to quantify latent evaluation dimensions from dissertation reviews. The framework integrates Shapley Additive exPlanations (SHAP) to ensure the interpretability of model predictions, combining deep semantic modeling with SHAP to quantify characteristic importance in academic evaluation. The experimental results demonstrate that the implemented model outperforms baseline methods in accuracy, precision, recall, and F1-score. Furthermore, its interpretability mechanism reveals key evaluation dimensions experts prioritize during the paper assessment. This analytical framework establishes an interpretable soft-sensor paradigm that bridges NLP with substantive review principles, providing actionable insights for enhancing dissertation improvement strategies.
Open AccessAn Interpretable Soft-Sensor Framework for Dissertation Peer Review Using BERT
Graduate education has entered the era of big data, and systematic analysis of dissertation evaluations has become crucial for quality monitoring. However, the complexity and subjectivity inherent in peer-review texts pose significant challenges for automated analysis. While natural language processing (NLP) offers potential solutions, most existing methods fail to adequately capture nuanced disciplinary criteria or provide interpretable inferences for educators. Inspired by soft-sensor, this study employs a BERT-based model enhanced with additional attention mechanisms to quantify latent evaluation dimensions from dissertation reviews. The framework integrates Shapley Additive exPlanations (SHAP) to ensure the interpretability of model predictions, combining deep semantic modeling with SHAP to quantify characteristic importance in academic evaluation. The experimental results demonstrate that the implemented model outperforms baseline methods in accuracy, precision, recall, and F1-score. Furthermore, its interpretability mechanism reveals key evaluation dimensions experts prioritize during the paper assessment. This analytical framework establishes an interpretable soft-sensor paradigm that bridges NLP with substantive review principles, providing actionable insights for enhancing dissertation improvement strategies.
- 8 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 11875-11891
Silent Intruders: Unveiling Security Flaws in BLE Smart Locks
Smart locks are increasingly deployed as critical consumer electronics in smart city environments, offering keyless access and remote management through Bluetooth Low Energy (BLE). Their widespread use across diverse spatial and temporal contexts introduces unique security challenges that remain underexplored. In this work, we conduct a comprehensive analysis of prior CVEs, academic studies, and technical reports to summarize seven established attack vectors. Additionally, we uncover three novel threatsCompromise of Shareable Access Credentials, SDK Exploitation, and Time-Reset Replay, exposing critical but previously overlooked flaws. Our comprehensive evaluation of 18 commercial BLE smart locks reveals that 14 devices remain vulnerable to attacks, affecting over 20 million users. These attacks enable unauthorized access, service denial, and bypass loggingeven on devices claiming advanced security. To mitigate such risks, we propose DGCSafe, a lightweight, OTA-updatable architecture tailored for constrained IoT devices. DGCSafe incorporates context-aware defenses and supports temporal state validation. Formal verification and real-world testing confirm its effectiveness. Our work provides actionable insights into securing smart lock deployments and demonstrates the importance of intelligent analytics for safeguarding consumer electronics in smart city infrastructures.
- 8 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 11875-11891
Silent Intruders: Unveiling Security Flaws in BLE Smart Locks
Smart locks are increasingly deployed as critical consumer electronics in smart city environments, offering keyless access and remote management through Bluetooth Low Energy (BLE). Their widespread use across diverse spatial and temporal contexts introduces unique security challenges that remain underexplored. In this work, we conduct a comprehensive analysis of prior CVEs, academic studies, and technical reports to summarize seven established attack vectors. Additionally, we uncover three novel threatsCompromise of Shareable Access Credentials, SDK Exploitation, and Time-Reset Replay, exposing critical but previously overlooked flaws. Our comprehensive evaluation of 18 commercial BLE smart locks reveals that 14 devices remain vulnerable to attacks, affecting over 20 million users. These attacks enable unauthorized access, service denial, and bypass loggingeven on devices claiming advanced security. To mitigate such risks, we propose DGCSafe, a lightweight, OTA-updatable architecture tailored for constrained IoT devices. DGCSafe incorporates context-aware defenses and supports temporal state validation. Formal verification and real-world testing confirm its effectiveness. Our work provides actionable insights into securing smart lock deployments and demonstrates the importance of intelligent analytics for safeguarding consumer electronics in smart city infrastructures.
- Open Access
Enhancing Long-Term Action Quality Assessment: A Dual-Modality Dataset and Causal Cross-Modal Framework for Trampoline Gymnastics
Action quality assessment (AQA) plays a pivotal role in intelligent sports analysis, aiding athlete training and refereeing decisions. However, existing datasets and methods are limited to short-term actions, lacking comprehensive spatiotemporal modeling for complex, long-duration sequences like those in trampoline gymnastics. To bridge this gap, we introduce Trampoline-AQA, a novel dataset comprising 206 video clips from major competitions (2018–2024), featuring dual-modality (RGB and optical flow) data and rich annotations. Leveraging this dataset, we propose a framework comprising a Temporal Feature Enhancer (TFE) and a forward-looking causal cross-modal attention (FCCA) module, which improves action quality assessment by delivering more accurate and robust scoring for long-duration, high-speed routines, particularly under motion ambiguities. Our approach achieves a Spearman correlation of 0.938 on Trampoline-AQA and 0.882 on UNLV-Dive, demonstrating superior performance and generalization capability.
Open AccessEnhancing Long-Term Action Quality Assessment: A Dual-Modality Dataset and Causal Cross-Modal Framework for Trampoline Gymnastics
Action quality assessment (AQA) plays a pivotal role in intelligent sports analysis, aiding athlete training and refereeing decisions. However, existing datasets and methods are limited to short-term actions, lacking comprehensive spatiotemporal modeling for complex, long-duration sequences like those in trampoline gymnastics. To bridge this gap, we introduce Trampoline-AQA, a novel dataset comprising 206 video clips from major competitions (2018–2024), featuring dual-modality (RGB and optical flow) data and rich annotations. Leveraging this dataset, we propose a framework comprising a Temporal Feature Enhancer (TFE) and a forward-looking causal cross-modal attention (FCCA) module, which improves action quality assessment by delivering more accurate and robust scoring for long-duration, high-speed routines, particularly under motion ambiguities. Our approach achieves a Spearman correlation of 0.938 on Trampoline-AQA and 0.882 on UNLV-Dive, demonstrating superior performance and generalization capability.