امروزه شبكه‌هاي جهاني وب تبديل به يكي از ابزارهاي مورد نياز بشر شده‌اند كه توسط كاربران بسياري در سراسر جهان مورد استفاده قرار مي‌گيرند. مساله‌اي كه در اين حوزه وجود دارد گستردگي بسيار زياد اينترنت و مطالب آن است كه اين گستردگي روز به روز و با سرعت بسيار زياد در حال افزايش است. اكنون يكي از مشكلاتي كه پديد مي‌آيد، اتلاف وقت كاربران براي دستيابي به كالا ها و خدمات مورد نياز آن‌ها است و ممكن است در بسياري مواقع كالا يا خدمت مورد نياز خود را پيدا نكنند، در نتيجه ارائه و پيشنهاد كالا يا خدمات مناسب به كاربران در زمينه هاي مختلف مطابق با نيازها آن‌ها امري بسيار حياتي محسوب مي گردد و يكي از روش هاي بسيار پركاربرد براي اين مساله استفاده از سيستم هاي توصيه‌گر مي‌باشد. ?سيستم‌هاي توصيه‌گر به منظورجلوگيري از اتلاف وقت كاربران، محصولاتي را به آنها پيشنهاد مي‌كنند كه به احتمال زياد مورد علاقه آن‌ها هستند و هنوز آن‌ها را نديده‌اند. الگوريتم پالايش گروهي به عنوان يكي از معروف ترين و پركاربرد ترين الگوريتم‌ها براي پياده سازي يك سيستم توصيه‌گر شناخته مي‌شود. سيستم‌هاي مبتني بر اين الگوريتم بر اساس سوابق جستجو و ابراز علاقه‌مندي كاربر به كالا‌ها و خدمات مختلف و با توجه اطلاعات دريافتي از ديگر كاربران با علاقه‌مندي‌هاي مشابه به كاربر هدف پيشنهادات جديدي را ارائه مي‌كنند. سيستم‌هاي توصيه‌گر اغلب با چالش مهمي به نام تنكي داده‌ها مواجه هستند، زيرا ماتريس‌هاي تعامل كاربر–اقلام در پياده‌سازي‌هاي واقعي معمولاً بيش از ??? تنك هستند. اين مسئله تأثير منفي بر دقت و كارايي توصيه‌ها دارد، به‌ويژه براي كاربران جديد و محصولات خاص يا كم‌تعامل. در حالي كه روش‌هاي موجود تلاش مي‌كنند با افزودن اطلاعات جانبي يا تغيير در طراحي سيستم اثر تنكي داده را كاهش دهند، اغلب مشكل اصلي يعني كمبود داده‌هاي تعاملي را ناديده مي‌گيرند. در اين پژوهش، يك چارچوب جديد معرفي مي‌شود كه از يك مدل ديفيوژني براي توليد امتيازدهي‌هاي مصنوعي با كيفيت بالا در سيستم‌هاي توصيه‌گر استفاده مي‌كند. به طور مشخص، از يك مدل ديفيوژني طراحي‌شده براي داده‌هاي جدولي استفاده مي‌كنيم تا توزيع مشترك و پيچيده‌ي سه‌تايي‌هاي كاربر–اقلام–امتياز را ياد بگيرد و سپس امتيازهاي مصنوعي توليد كند كه از نظر آماري سازگار هستند. براي تضمين كيفيت امتيازهاي توليدشده، يك روش نوآورانه براي شناسايي نويز بر اساس تحليل الگوهاي رفتاري پيشنهاد مي‌كنيم. اين روش امتيازهايي را كه با ترجيحات كاربران و ويژگي‌هاي آيتم‌ها همخواني ندارند، شناسايي كرده و حذف مي‌كند. براي ارزيابي كارايي چارچوب پيشنهادي، از مدل‌هاي پالايش گروهي سنتي و پالايش گروهي مبتني بر شبكه‌هاي عصبي استفاده شده است. آزمايش‌ها روي دو مجموعه‌داده واقعي با سطوح مختلف تنكي داده (با نگهداشت داده از 5% تا ???%) انجام شد و بهبودهاي قابل‌توجهي را نشان داد. به طور خاص، پالايش گروهي سنتي مي‌تواند خطاي RMSE را تا ??% كاهش دهد و نيز پوشش امتيازدهي را تا ?? % افزايش دهد. همچنين، پالايش گروهي عصبي پاسخ‌هاي دقيق‌تر و ظريف‌تري ارائه مي‌دهد و زماني بيشترين كارايي را دارد كه نسبت افزايش داده پايين باشد. اين موضوع نشان مي‌دهد كه مدل‌هاي مختلف به انواع متفاوتي از افزايش داده نياز دارند.   

The primary aim of thisresearch is to design and implement an intelligent and efficient framework forintrusion detection in Internet of Things (IoT) devices using a novel FederatedLearning (FL) approach. With the rapid growth of IoT applications in domainssuch as healthcare, industry, agriculture, and smart cities, a massive amountof data is generated by connected devices. Ensuring the security and privacy ofthis data has become a critical challenge. Traditional centralized intrusiondetection systems (IDSs) are no longer suitable due to their high communicationoverhead, limited device resources, and hardware heterogeneity. To address these challenges,this thesis introduces a new framework called ASA (Adaptive Smart Agent). ASAemploys an adaptive agent layer that monitors device resources and dynamicallyclusters IoT devices based on their computational power, memory capacity, andbandwidth. For each cluster, an appropriately scaled learning model isassigned. The training process is performed locally on devices, and only modelupdates are transmitted to the central server, thereby reducing communicationcosts and preserving user privacy.Experimentalevaluations on benchmark IoT datasets demonstrate that ASA significantlyoutperforms conventional FL-based methods in terms of detection accuracy,communication efficiency, and participation fairness. It effectively mitigatescritical issues such as device dropouts, non-IID data distribution, and networkinstability, while maintaining robustness and stability in heterogeneousenvironments.Theresults highlight that the proposed ASA framework enhances the accuracy andscalability of IoT intrusion detection systems while ensuringprivacy-preserving distributed learning. Future work can focus on acceleratingmodel convergence, improving fault tolerance, and integrating ASA with edge andfog computing infrastructures to enable real-world deployment in large-scaleIoT ecosystems.  

   In this research, a hybrid algorithm called WOA-Q Learning is proposed for real-time task scheduling in Fog-Cloud environments. The algorithm integrates the global exploration capability of the Whale Optimization Algorithm (WOA) with the adaptive decision-making mechanism of Q-Learning, aiming to optimize resource allocation and minimize delay. Simulations conducted in MATLAB across scenarios with 25 to 100 tasks demonstrated that the proposed method outperforms benchmark algorithms such as EDF, PSO, WOA, and QL in four key performance metrics: total delay, energy consumption, deadline miss ratio, and scheduling efficiency. The WOA-Q algorithm achieved up to 25% reduction in total delay, 20% lower energy consumption, and a deadline miss ratio of around 0.05. Although its execution time is slightly higher due to computational complexity, the overall performance improvement justifies the trade-off. The results confirm that combining metaheuristic and reinforcement learning techniques provides an effective and intelligent approach for real-time scheduling, with significant potential for applications in IoT, edge computing, and industrial control systems.          Keywords: Optimization, Real-time scheduling, Cloud-fog environment, Internet of Things      

  Abstract Multiple sclerosis (MS), as a chronic and disabling disease of the central nervous system, has created many challenges in the field of diagnosis and treatment for doctors and health systems. Rapid and accurate identification of lesions caused by this disease in MRI images due to structural similarities with other brain tissues requires the use of accurate and advanced image processing and machine learning methods. In this study, an optimized model called DE-CNN-Gray is presented for automatic diagnosis of MS from gray-scale MRI images. In this model, a convolutional neural network is first designed and then the network structure including the number of layers and effective parameters is optimized using the Differential Evolution algorithm. The main goal of this optimization was to increase the classification accuracy and reduce the computational complexity of the model. Model evaluation using 5-Fold validation showed that the proposed method performed very well in identifying MS patients with an accuracy of 99.40%, sensitivity of 98.89%, positive accuracy of 99.90%, and F1 score of 99.33%. The results show that the DE-CNN-Gray method, using gray images and meta-heuristic algorithms, can be used as an accurate, fast, and low-cost tool for developing MS diagnosis systems and play an effective role in improving the treatment process and reducing treatment costs . Keywords: MS, Deep Learning, Convolutional Neural Network, Differential Evolution Algorithm, MRI, DE-CNN-Gray,Medical Diagnosis, Image Processing.

   With the rapid development of smart cities and the growing need for accurate and real-time analysis of road infrastructure, the design of AI-based systems capable of perceiving, analyzing, and recording environmental data has become increasingly crucial. In this regard, the present study focuses on the design and implementation of an innovative system for the automatic detection, tracking, and localization of traffic signs. This system not only pushes technical boundaries but also makes a significant contribution to the localization of traffic-related data. The advanced YOLOv9 model is employed for precise traffic sign detection, while the powerful ByteTrack algorithm ensures continuous tracking. What truly distinguishes this research is the novel application of federated learning using the FedAvg algorithm—implemented for the first time in the domain of traffic sign recognition. This method enables the training of models on heterogeneous datasets, including two distinct subsets, DFG and Mapillary, without requiring physical data aggregation. This approach not only preserves data privacy but also significantly enhances the generalization capability of the model. On the data side, the study introduces a rich and unprecedented dataset comprising 14,111 images and over 19,000 traffic sign instances across 118 classes. The data was collected over two years in varying temporal conditions (morning, noon, evening, night) and all four seasons, spanning urban, rural, and interurban areas across the country using mobile phone cameras. The images were meticulously annotated using the Makesense tool in both YOLO (.txt) and Pascal VOC (.xml) formats. The system’s performance, evaluated through 6-Fold Cross Validation, demonstrates its high accuracy, achieving a remarkable mAP50 of 95.66%. This not only reflects the model's robustness in real-world conditions but also shows a clear advantage over various versions of YOLO from v5 to v11. The initial idea for this project stemmed from a proposal by Tehran Municipality to design a digital map of traffic signs. However, the outcomes of this research go far beyond a municipal application and offer valuable tools for navigation systems, intelligent vehicles, spatial analytics, and the development of a national traffic sign map in Iran. This work presents a seamless integration of cutting-edge technology, locally-driven data, and modern AI architectures to pave the way for a smarter and safer future on the country’s roads.

includes 2605 training samples and 1321 test samples. The labeling of these data was done with two classes positive and negative bythree annotators using the majority voting method. In this study, five different architectures, namely Bert-fa-zwnj-base, Bert-fa-base-uncased, LSTM, GRU, and Distil-bert, were employed for sentiment analysis, and these models were evaluated with two optimizers, SGD and Adam. The results indicate that the Bert-fa-base-uncased model performed the best on both datasets, achieving an accuracy of 93% on the Twitter dataset and 80% on the Instagram dataset. Furthermore, the Adam optimizer outperformed SGD. This research demonstrates that the use of deep learning-based models, especially Bert-fa-base-uncased, can effectively perform sentiment analysis on Persian texts with high accuracy and efficiency, processing data generated on widely used platforms such as Instagram and Twitter effectively.  

Sleep is one of the fundamental human needs that significantly impacts physical and mental health. Stress during sleep can lead to sleep disorders and related health issues, making accurate prediction of sleep stress particularly important. This thesis explores the detection of sleep stress using deep learning, specifically focusing on LSTM and GRU recurrent neural network models, as well as a hybrid model combining the two. The aim of this research is to provide an efficient and accurate model for predicting and detecting sleep stress based on the SaYoPillow dataset. We used 10-fold Cross Validation. Various models were evaluated, and the results showed that the hybrid Bidirectional LSTM-GRU model achieved the best performance with an accuracy of 1.00, precision of 1.00, recall of 1.00, and an F1 score of 1.00, outperforming individual LSTM and GRU, and MLP models in detecting all 5 levels of sleep stress. The use of a confusion matrix and evaluation metrics such as accuracy, precision, recall, and F1 score demonstrated that the hybrid model not only has high accuracy in detecting positive cases but also reduces errors related to identifying negative cases. This research highlights that deep learning models, particularly the hybrid Bidirectional LSTM-GRU model, can be effective tools for detecting sleep stress, thereby contributing to improved sleep quality and overall health. The development of these models can assist healthcare professionals in providing appropriate preventive and therapeutic strategies for managing sleep stress.

Abstract Hydrogen production as a clean and sustainable energy source, particularly through water splitting, is considered one of the fundamental challenges in the field of renewable energy. In this study, the process of hydrogen production through photocatalytic water splitting using photocatalysts Bi?WO? and CuBi?O? with different weight ratios is investigated and analyzed. The two photocatalysts, CuBi?O? and Bi?WO?, were synthesized via a hydrothermal method and subsequently combined in various weight percentages. Due to their unique properties in light absorption and charge transfer enhancement, these photocatalysts have the potential to exhibit high efficiency in hydrogen production from water. To characterize the Bi?WO? and CuBi?O? photocatalysts, several analytical techniques were employed. Subsequently, the impact of forming a heterogeneous junction between these two materials on water splitting performance was examined. All reactions were conducted under UV-Visible light in a 160 mL quartz reactor. Experimental results indicate that although the pure Bi?WO? and CuBi?O? photocatalysts produce hydrogen at rates of 131.87 ?molg?¹ h?¹ and 165.56 ?molg?¹ h?¹, respectively, the heterogeneous Bi?WO?/CuBi?O? junction in the optimized sample significantly increases the hydrogen production rate to 341.25 ?molg?¹ h?¹ compared to the individual photocatalysts. This enhancement in efficiency is attributed to improved light absorption, increased electron and hole lifetimes, and reduced electron recombination.   Keywords: Photocatalyst, CuBi?O?, Bi?WO?, Water Splitting, Heterojunction

   Abstract Digitalization in the field of government and then land administration as a government service, has come to the integration of processes and data through a single Point of Access called the single-window system. However, assessing the maturity level of this system and tracking its progress remains a major challenge. This thesis aims to propose a descriptive maturity model for evaluating the land administration single-window system, also known as Iranland . The proposed model is developed through a literature review and the integration of reference methodologies, frameworks and models such as the World Trade Organization’s Single Window Assessment Methodology (SWAM), the World Customs Organization's (WCO) Single Window Maturity Model, the Capability Maturity Model Integration (CMMI), the Open Group Architecture Framework (TOGAF), the United Nations e-Government Development Index (EGDI), the Organization for Economic Cooperation and Development (OECD) maturity frameworks, Iran’s five-level e-Government maturity model, and several other models and frameworks. In this proposed maturity model, the maturity of the system is assessed based on several key process areas, including integration level, data management, user interface, stakeholder participation, tra  arency, system performance, smart monitoring and control, and more. The proposed maturity model is presented in five maturity levels (Initial, Standard, Integrated, Advanced, and Optimization & Innovation), which align with phases A, B, and C of the TOGAF architecture development method, CMMI integration, and Iran’s five-level e-Government maturity model. The model was evaluated through comparative analysis with its reference models. The results indicate that the key benefits of the proposed maturity model are the integration of criteria from several recognized models, covering diverse indicators, and its compatibility with the specific characteristics of the land management single-window system. However, the model also has limitations, including the lack of a precise method for assessing weight to indicators, limited coverage of advanced security considerations, and dependency on data quality and availability, which could be explored in future research. Keywords: Maturity model, land administration single window system, single window, e-Government, maturity assessment.

   Common signals recorded from neural activity in the brain, such as EEG or MEG, are non-stationary signals. Therefore, time-frequency analysis tools are well suited to analyze such signals. Over the years, time-frequency analysis has found numerous applications in computational neuroscience, from diagnosing diseases such as schizophrenia in adults to understanding autism spectrum disorders in children and identifying seizures in infants. With advances in computational resources and various types of imaging tools, time-frequency analysis is on the verge of becoming one of the most important mainstream tools in the field of computational neuroscience. We want to examine the methods of examining the properties of brain networks using time-frequency representation tools. In this thesis, time-frequency based phase synchronization methods are introduced and analyzed. Time-frequency distribution includes WVD, Rihaczek, spectrogram, EMBD, ... methods, which are briefly reviewed. Using time-frequency representations, we can access the time-varying properties of brain networks. The results presented in this thesis show that time-frequency-based phase synchrony is a good measure for examining the effective connections of brain networks, which allows us to use powerful analytical tools to investigate and better understand brain networks and their dynamic behavior over time. The results of this evaluation show that phase synchronization methods based on time-frequency representation can have applications in analyzing brain networks, including studying brain signals of infants and adults, diagnosing diseases such as epilepsy, and studying and helping to diagnose diseases such as Parkinson's and autism, which can lead to better and more effective treatment