Adaptive processing of infrared imagery with a wide dynamic range is a formidable task in thermal imaging systems. Existing work has mainly addressed histogram equalization and statistical transforms, which fail to maintain details for extreme temperature swings in the same scene. Standard techniques lead to a loss of detail in high-temperature areas or a lack of enough contrast in low-temperature environments, which keeps infrared systems ineffective in challenging situations. Research overcomes such limitations by formulating a Wide Dynamic Range processing algorithm based on Double Deep Q-Network (WDR-DDQN), an intelligent framework capable of optimizing image enhancement parameters according to scene properties. WDR-DDQN applies a double deep Q-network (DDQN) reinforcement learning structure with experience replay and prioritization, dueling networks, and multi-scale thermal decomposition methods that enhance adaptive local and global contrast, maintaining important thermal details. The WDR-DDQN framework attained a 42% improvement in structural similarity index, a 37% improvement in target recognition rate, and a 28% improvement in processing latency over conventional methods on various thermal datasets. These gains allow for much improved visualization and detection performance in difficult cases with extreme temperature gradients. Research promotes deep reinforcement learning in real-world infrared image processing applications, thereby providing novel capabilities for security systems, self-driving cars, and military usage where accurate thermal imaging over a broad range of temperatures is crucial.
Large Language Models (LLMs) have demonstrated strong capabilities in natural language generation; however, they remain prone to hallucinations and often lack verifiable grounding when addressing knowledge-intensive queries. Retrieval-Augmented Generation (RAG) alleviates this limitation by incorporating external knowledge sources, yet most existing approaches apply retrieval indiscriminately, which can introduce irrelevant context and increase computational overhead. In this paper, we propose Auto-RAG, an adaptive retrieval-augmented generation framework with conditional routing that dynamically determines whether external knowledge retrieval is necessary for a given query. The framework employs a lightweight query analysis module to assess query complexity, intent, and lexical specificity, enabling intelligent routing between a direct LLM generation pathway and a retrieval-augmented reasoning pipeline. To enhance retrieval effectiveness, the proposed approach combines semantic embedding-based search with keyword-based BM25 through a unified fusion and re-ranking mechanism, ensuring high-quality and contextually relevant evidence selection. In addition, we introduce a trust-aware evaluation framework that jointly measures answer quality, faithfulness to retrieved evidence, hallucination rate, and retrieval efficiency. Experimental results on a nutrition knowledge corpus derived from Human Nutrition demonstrate that the proposed framework achieves a strong balance between fluency and factual reliability. Compared to conventional always-on retrieval strategies, Auto-RAG reduces retrieval usage by approximately 30% while maintaining competitive answer quality and evidence-grounded performance. The results also show that the effectiveness of evidence filtering depends on its configuration, with overly aggressive filtering potentially removing useful supporting evidence. While LLM-only responses may exhibit higher linguistic fluency, they lack grounding guarantees, underscoring the importance of adaptive retrieval in safety-critical domains. Overall, this work presents an efficient and reliable approach for domain-specific knowledge access and highlights conditional retrieval as a scalable paradigm for improving trustworthiness in large-scale language systems.
The rapid growth of smart devices has increased the demand for robust and scalable encryption mechanisms while expanding the attack surface to cyber threats, especially in Internet of Things (IoT) environments. Recent advances in Deep Learning (DL) have created new opportunities to improve encryption methods and enable secure data storage and transmission against emerging cyber threats. This paper presents an optimized security framework that integrates DL with an optimized cryptographic scheme to provide secure and effective data protection. A Hybrid Convolutional Neural Network and Long Short-Term Memory (HCL) model is used to generate candidate cryptographic keys, then Snake Optimizer (SO) is applied to select an optimal key that maximizes entropy and minimizes encryption and decryption time. To enhance data confidentiality, Advanced Encryption Standard and Blowfish (ABF) algorithms are used. In key selection experiments, the results showed that HCL-SO achieved the lowest average key selection time (211 ms), with improvements of 17.51%, 9.14%, and 16.55%, compared to the HCL integrated with Grey Wolf Optimization (GWO), Honey Badger Algorithm (HBA), and Harris Hawks Optimization (HHO), respectively. For data confidentiality, the proposed key generation mechanism with the ABF algorithm achieved an average of 64.2% lower encryption latency and 25.3% higher throughput compared to the standard AES-Blowfish encryption model. Furthermore, the security of the proposed method was validated using NIST statistical tests and key sensitivity analysis. The results demonstrated high randomness and strong sensitivity to small changes, indicating that the proposed framework provides an effective and secure approach for protecting IoT data applications.
Vehicular Ad hoc Networks are highly dynamic wireless networks that enable communication between vehicles and roadside infrastructure. They enhance road safety by supporting driver-assistance services and advanced transportation applications, such as automated driving and intelligent transportation systems. These networks are characterized by dynamic topology, intermittent connectivity, and limited resources. Most existing blockchain-based authentication approaches rely on roadside units or continuous Internet connectivity, which increases deployment costs and limits vehicle autonomy in areas without network coverage. This article proposes an authentication mechanism that combines a consortium blockchain with integrated edge computing to enable vehicle-to-vehicle authentication without roadside units and without continuous Internet connectivity. Each vehicle is equipped with an Edge Box that performs local data processing, real-time inter-vehicle authentication, and secure communication. The system uses an elliptic-curve–based key exchange method and symmetric encryption to ensure communication confidentiality and integrity. Experiments conducted on two Raspberry Pi-based Edge Boxes over 10 trials show a communication cost of 64 bytes per authentication exchange, reducing communication overhead by 54%–97% compared to existing blockchain-based schemes, with an average mutual-authentication delay of 58.46 ms.
Math word problem (MWP) is a famous type of math problems that require natural language understanding along with mathematical reasoning. MWP solvers should understand the problem’s context, identify relevant quantities, infer relationships and finally deduce the solving equation. In order to develop such solvers, suitable MWP dataset should be constructed to be utilized in training these solvers. Despite the presence of MWP datasets in different languages, few reliable Arabic MWP datasets were publicly found. Addressing this gap, we studied the most recent data augmentation strategies and their adaptations to the Arabic language and applied them to the most promising Arabic MWPs “ArMATH” extending its vocabulary and generalization capabilities. Our experiments demonstrated that augmenting the training data helps the models better solve MWPs, increasing the equation and answer accuracies by 5.2 and 5.1 scoring 73 and 74.2 respectively. ArMATH + + , reliable Arabic MWPs dataset was introduced. With a size of 33,918 samples, ArMATH + + can be utilized in training deep neural networks and LLMs finetuning.
