Abstract The rapid advancement of generative artificial intelligence has significantly increased the realism and accessibility of deepfake content, posing critical challenges to digital media authenticity and security. This study proposes a hybrid deepfake detection framework that integrates motion-driven analysis with temporal sequence modeling to improve detection robustness. The approach combines frame-differencing techniques to capture pixel-level motion inconsistencies with Long Short-Term Memory (LSTM) networks to model temporal... The rapid advancement of generative artificial intelligence has significantly increased the realism and accessibility of deepfake content, posing critical challenges to digital media authenticity and security. This study proposes a hybrid deepfake detection framework that integrates motion-driven analysis with temporal sequence modeling to improve detection robustness. The approach combines frame-differencing techniques to capture pixel-level motion inconsistencies with Long Short-Term Memory (LSTM) networks to model temporal dependencies across facial video sequences. Experiments were conducted on preprocessed facial datasets derived from FaceForensics++ and DeepFakeDetection, ensuring standardized inputs through face extraction, alignment, and quality filtering. Motion intensity was computed from frame to frame differences, enabling the effective identification of temporal anomalies via statistical thresholding. Qualitative analysis using temporal montage visualization demonstrated strong motion coherence in real sequences and detectable inconsistencies in manipulated content. The LSTM-based temporal modeling further enhanced the system by capturing motion trends, activity transitions, and sequence level patterns. Results indicate that the proposed framework effectively identifies high motion anomalies and improves interpretability through temporal smoothing. However, evaluation using pseudo ground truth revealed limitations in distinguishing low and moderate activity classes, highlighting the constraints of motion only features. Overall, the study demonstrates that integrating spatial motion cues with temporal learning significantly enhances deepfake detection performance. Future work will focus on incorporating spatio temporal deep learning models and multimodal features to further improve classification accuracy and robustness. Read More Read Less

Abstract This study presents a comparative analysis of meta-learning-driven approaches for few-shot font adaptation, addressing the critical challenge of recognizing unseen font styles with limited labeled samples. As data scarcity and domain shifts hinder conventional deep learning models, we investigate three adaptation paradigms: baseline fine tuning, metric based Prototypical Networks, and Model-Agnostic Meta-Learning (MAML). Through rigorous evaluation across 1-shot, 5-shot, and 10-shot settings, we assess... This study presents a comparative analysis of meta-learning-driven approaches for few-shot font adaptation, addressing the critical challenge of recognizing unseen font styles with limited labeled samples. As data scarcity and domain shifts hinder conventional deep learning models, we investigate three adaptation paradigms: baseline fine tuning, metric based Prototypical Networks, and Model-Agnostic Meta-Learning (MAML). Through rigorous evaluation across 1-shot, 5-shot, and 10-shot settings, we assess classification accuracy, data efficiency, adaptation velocity, and knowledge retention. Empirical results demonstrate that MAML consistently achieves superior performance, attaining 50.9%, 58.3%, and 62.5% accuracy, respectively, while exhibiting the fastest convergence and lowest variance across target fonts. Feature representation analysis via t-SNE visualizations confirms that meta-learning facilitates effective domain alignment, transforming scattered target samples into compact, class-discriminative clusters. Furthermore, confusion matrix analysis reveals reduced inter-class confusion, particularly for structurally similar digits. Crucially, our forgetting measure indicates that MAML preserves source-domain knowledge with near-zero catastrophic forgetting, unlike conventional fine-tuning which suffers significant performance degradation. These findings establish meta learning as a principled framework that balances rapid adaptation, data efficiency, and long term stability. The proposed approach offers scalable solutions for real world applications that require continuous adaptation to novel visual styles under data constrained regimes, advancing fewshot learning methodologies in font recognition and broader low resource visual recognition tasks. Read More Read Less

Abstract Preventing traffic accidents in intelligent connected vehicles requires accurate anomaly detection and behavioral modeling. Traditional rule-based systems are insufficient for dynamic traffic environments. This study presents an integrated architecture for vehicle trajectory anomaly detection and scenario-based temporal simulation. The framework synthesises deep learning spatiotemporal modelling, driver fatigue monitoring, and data driven scenario generation into a unified six layer system. Utilizing a dataset of 5,000 samples... Preventing traffic accidents in intelligent connected vehicles requires accurate anomaly detection and behavioral modeling. Traditional rule-based systems are insufficient for dynamic traffic environments. This study presents an integrated architecture for vehicle trajectory anomaly detection and scenario-based temporal simulation. The framework synthesises deep learning spatiotemporal modelling, driver fatigue monitoring, and data driven scenario generation into a unified six layer system. Utilizing a dataset of 5,000 samples with 19 features, the core analysis module combines reconstruction based models (DAGMM, VAE) and prediction-based networks (STSSN, STCL) for anomaly detection. A supervised classifier achieves 98.24% accuracy in behavior classification, identifying risk probability and obstacle distance as dominant predictors. A distinguishing feature is the Scenario-Based Temporal Simulation Engine, which models behavioral evolution using probabilistic transition matrices (e.g., follow ï‚® yield ï‚® stop). This engine enables continuous state evolution and closed loop simulation, capturing hierarchical safety responses under varying conditions such as congestion or approaching obstacles. Specific scenarios include approaching obstacles, visibility degradation, and lane maneuvers under traffic pressure. While temporal dynamics are simulated due to dataset constraints, the system provides interpretable decision rules and risk assessments. Results indicate driving behavior emerges from threshold-based interactions rather than independent features. Despite limitations regarding real-time trajectory data, this holistic, risk-centric approach significantly advances autonomous driving safety by enabling proactive risk assessment and robust decision support for nextgeneration intelligent transportation systems. Future work should incorporate real time timestamped datasets to enhance generalization. Read More Read Less