Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 22 4 2026 3 1 KANFlow: A Novel Approach to Encrypted Traffic Identification Using Kolmogorov-Arnold Network 18 3 FA ali rahnema Ph.D. student of IT, University of Qom, Qom, Iran zahra akhoodad Faculty Member, Research Center for Development of Advanced Technology, Tehran, Iran With the growing usage of encryption protocols like VPN, and Tor in digital communication, identification and classification of encrypted traffic has been one of the core issues in network security and traffic management. It is a major contributor to quality of service (QoS) assurance, resource allocation, user identification, and anomaly detection. But the sophistication of encrypted traffic structure and the vagueness of behavioral patterns have drastically decreased the effectiveness of conventional approaches like deep packet inspection (DPI). In spite of the progress, typical deep learning models also encounter great difficulties in dealing with encrypted data; they typically need a huge amount of labeled data and lack the capacity to analyze unbalanced data. To tackle these difficulties, this study proposes a novel hybrid architecture named seqKAN with enhanced interpretability and high accuracy. seqKAN integrates the temporal modeling capability of sequential networks like LSTM with the distinctive characteristics of Kolmogorov-Arnold networks (KAN), such as examining nonlinear relationships and intrinsic mathematical transparency. The framework also enjoys high flexibility and generalizability with the use of modules like Reproducible Hilbert Space Mapping (RKHS) and Neural Ordinary Differential Equations (ODE). Experiments are performed on benchmark datasets comprising Tor and VPN traffic (ISCXTor2016 and ISCXVPN2016). In this context, by meticulously filtering out the streams and addressing unbalanced data via class weighting, the model's stable performance is guaranteed. Ablation Study demonstrate that the inclusion of the RKHS layer significantly contributes to the improvement of the model's accuracy and robustness, particularly in encrypted settings. Among the models compared, the seqKAN approach delivered the best performance in F1 score and demonstrated clear superiority in the classification of encrypted traffic. In addition, the interpretability of the model was quantitatively and qualitatively demonstrated with standard feature importance analysis techniques (SHAP and LIME) and KAN's inherent visual analysis. seqKAN successfully automatically extracted key features and patterns in the flow packets and clearly explained each decision; this transparency evidently illustrates the model's superiority over typical methods. Finally, this research shows that the seqKAN architecture provides a comprehensive, efficient, and reliable solution for intelligent traffic analysis in complex network environments by creating a smart balance between accuracy, computational efficiency, and interpretability. The findings of this research highlight the high potential of KAN-based hybrid models as the basis for the next generation of transparent and reliable network security tools

Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 22 4 2026 3 1 ADHD recognition by processing nonlinear features of ABR based on a new innovative method for extracting Geometry features from phase space trajectory 38 19 FA Ali Akbar Alidoust Ghadikolayi PhD Candidate in Biomedical Engineering, Azad University, South Branch, Tehran, Iran Ali Motie Nasrabadi Professor, Department of Biomedical Engineering, Shahed University, Tehran, Iran Saeed Malayeri Part-time Lecturer, Iran University of Medical Sciences, Tehran, Iran ADHD recognition in first years after birth is essential to consider a better treatment plan for patients and consequently helps children who suffer from this disorder to promote their communication abilities. According to the latest researches, ADHD recognition through ABR is a useful method for this purpose instead of surveys or other oral methods. Considering the basis of ABR which is nonlinear, so nonlinear processing methods can be more effective for classification normal and ADHD groups from each other than other linear classifications. In this paper ABR signal of two groups of children including 37 normal and 31 with ADHD disorder have been recorded in a rehabilitation center. To that end, in this study firstly by using embedding method, the two-dimensional ABR signal converted to three-dimensional signals to be ready for drawing and calculating the phase space arguments. In the phase space the accuracy of this hypothesis was checked that if ADHD and normal signals in different time would show different behavior in occupying and movement through the voxels in the phase space, for this purpose at first the phase space data should be produced, but it is important to have access to the raw data of phase space because if the methods that only print the phase space were been used the access of raw data would be impossible, Actually using these method for drawing phase space is in conflict with this study’s objectives, because it is not important to just draw the phase space of normal or ADHD signals but it is necessary to divide the phase space of individual signals to equal voxels and then determine that any sample of trajectory points is located in which of these voxels over the time. So, it is crucial to access the raw data of each axis of phase space of any signal to use them for diving that volume to voxels and then trace the attendance of trajectory in these voxels and make a new mapped signal that introduce the voxel number of each point of trajectory against time.  Therefore the source data of any axis of any signal has been produced by embedding method that it is discussed before, it means that we replace any x(t) by {x(t), x(t+lag), x(t+2 lag)} and they are three axis of new three dimensional space, so we can use these axis for diving this space to different voxels. The phase space should be divided to equal voxels depending on the length of each signal, then a novel method of classification has been developed by extracting new geometry features in the phase space. This approach includes dividing three-dimensional phase space to equal voxels and checking space voxel’s occupation by trajectory points in these voxels. As trajectory points has the same time sequence as time series of original signals so trajectories can be used in mentioned new method to check time occupation of trajectories in phase space. Then by changing each sample value by its voxel number of that sample, a mapping method was developed, this new mapped signal is the basis of later analysis and feature extractions in this study. Briefly, this mapped signal has been formed by taking some steps including producing the phase space signal from ABR, dividing the phase space to voxels and predicating voxel numbers to each sample. Then to find some distinctions between ADHD and normal group mapped signal, four groups of features have been developed in this study including temporal features, extremum features, histogram features, spatial occupational, Lyapunov features. Temporal features are some common features including min, max, mean, median, variance, skewness, kurtosis. But further in accuracy analysis of project it has been cleared that time or sample limited min and max (called extremum features) are more effective and can make a better distinction result. Histogram features focus is on the most repeated voxel numbers which means which voxels has been most occupied by trajectory during its journey through the phased space. Spatial occupational features deal with the span or extension of total trajectory path by introducing some features like the biggest or smallest voxel number which has been occupied by at least one point of trajectory, as it is can be interpreted from these two latter features, the subtraction of the biggest and smallest occupied voxel number can represent a vision of extension of trajectory. Another feature in this category is the total voxels quantity which are occupied by at least one of the points of the trajectory. The Lyapunov related feature also calculate the total number of non-zero voxels in a specific duration of time than can represent the chaotic grade of the signal, beside the difference of two adjacent time duration Lyapunov feature and show the gradient of chaotic level of the signal that shows the Lyapunov gradient. Finally, efficiency of this method and features have been evaluated and best result observed in local minimum feature extracted by the new method and using KNN and SVM classifiers, the best accuracy is about 98.53 that shows a significant increase in accuracy in comparison with linear processing approaches. In the other words, this distinction gained from the minimum feature shows that there are some places among the lower voxel numbers in trajectories which Normal group of signals desire to occupy more than ADHD group. According to anatomical data which is extractable from ABR signal in different points of time is reachable in the final mapped data also, because the time sequence of first samples is saved in the final data. It means that we can check the distinction areas in phase space and adopt them with time or their anatomical generating source. From the physiologic point of view, these distinction areas in the phase space are correspond to the activity of the primary auditory neurons in the cochlear nerve and lower levels of the brainstem.

Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 22 4 2026 3 1 Unsupervised Semantic Segmentation of RGB-D Images Using Combination of Conditional Random Field with Graph Cuts 52 39 FA Seyedsaeid Mirkamali Assistant Professor Department of Computer Engineering and IT, Payame Noor University, Tehran, Iran Semantic segmentation seeks to give a set of pixels depicting an object in an image suitable labels depending on their appearance and semantic characteristics. Though it is still one of the most difficult issues in image processing and computer vision, this work has attracted a lot of interest recently. The availability of RGBD sensors has introduced new possibilities for segmentation by incorporating depth information alongside color. However, effectively combining these modalities presents challenges due to misalignments and depth inaccuracies. This paper proposes CRFCut, a novel unsupervised segmentation method that utilizes a Conditional Random Field (CRF) model optimized with graph cuts to segment RGBD images into coherent regions. The method recursively divides regions into foreground and background layers, employing superpixel-based appearance segmentation for the RGB component and integrating depth cues to refine results. This approach enables robust segmentation, even in the presence of noisy or incomplete depth information. The CRFCut algorithm begins by separating the depth image into foreground and background regions using a median depth threshold. This initial step requires no preprocessing and provides the basis for further segmentation. Simultaneously, the RGB image is segmented into superpixels using an appearance-based approach, such as the mean-shift algorithm. These superpixels and the depth regions are combined within a CRF model, where labels are assigned by minimizing the energy function using the graph-cut α-expansion algorithm. The algorithm is applied recursively to subdivided regions, allowing finer segmentation in a parallelizable manner. The proposed method was evaluated on two datasets: the NYUv2 dataset and the MIT dataset. The NYUv2 dataset, which includes 1449 RGBD images with annotated object classes, demonstrated the superior performance of CRFCut compared to five state-of-the-art segmentation techniques in Table 1. In the MIT dataset, which provides human-labeled sequences of indoor and outdoor scenes, CRFCut achieved comparable or better results, even with depth maps generated from 2D images using existing estimation methods (Table 2). The RandIndex metric was used to evaluate segmentation accuracy, and qualitative results, as shown in Figures 3 and 4, highlight CRFCut’s  robustness, particularly with noisy or imprecise depth data. In summary, CRFCut introduces an unsupervised CRF-based approach that integrates RGB and depth information for accurate scene segmentation. By leveraging graph-cut optimization and a recursive structure, the method achieves high-quality segmentation results with minimal preprocessing. Despite some limitations, such as challenges in distinguishing adjacent objects with similar features, CRFCut offers a promising framework for real-time segmentation of RGBD images. Future work will address these limitations by incorporating supervised techniques and improving depth data quality for enhanced performance.  

Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 22 4 2026 3 1 A Deep Learning Based Method for Android Malware Detection 70 53 FA Ali Olyaei Torqabeh PhD Student of Software Engineering, Department of Computer Engineering, Faculty of Engineering, Ferdowsi University of Mashhad, Mashhad, Iran Abbas Rasoolzadegan Associate Professor of Software Engineering, Department of Computer Engineering, Faculty of Engineering, Ferdowsi University of Mashhad, Mashhad, Iran The Android operating system, an open-source platform supported by Google, has become a cornerstone of modern technology due to its widespread adoption in diverse devices, including smartphones, smart TVs, and wearables. This extensive reach has established Android as a dominant force in the global market but simultaneously made it a primary target for malware developers. The growing sophistication and frequency of mobile malware attacks pose significant challenges for users and Android app distribution platforms. These attacks exploit the open nature of the Android ecosystem and increasingly employ advanced techniques such as obfuscation, rendering traditional detection methods less effective. In response to these challenges, this study introduces an innovative approach to malware detection leveraging image and audio processing in combination with deep learning techniques. Our proposed methodology addresses the limitations of existing methods by providing a scalable, high-accuracy solution suitable for industrial deployment. The research is based on static analysis. During the static analysis, executable file bytes are transformed into audio signals, and features extracted from these signals are used to train a deep learning model. This model achieved an impressive accuracy of 99.3%, with a precision of 99.8% and a recall of 99.1%. The novelty of our approach lies in its ability to detect obfuscated malware, a critical and challenging aspect of modern malware detection. By mapping executable files to the audio domain in static analysis, our method effectively reduces computational complexity while enhancing detection accuracy. The proposed framework was validated on a diverse and comprehensive dataset, showcasing its capability to distinguish between benign and malicious applications with high reliability. Furthermore, the method's design ensures practical applicability in real-world scenarios, particularly in app distribution platforms where rapid and accurate malware detection is crucial. This research contributes a novel, efficient, and scalable malware detection solution that addresses the challenges posed by obfuscation and computational demands. The proposed framework not only advances the state-of-the-art in Android malware detection but also lays the groundwork for future research exploring hybrid analysis techniques and real-time detection capabilities.  

Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 22 4 2026 3 1 Presenting a model for establishing trust in inter-vehicle networks based on blockchain using fuzzy inference and Chord structure 100 71 FA niloufar khosravirad Ph.D. Candidate, Department of Computer Engineering, Islamic Azad University, Qom Branch, Qom, Iran reza ahsan Assistant Professor, Faculty of Computer Engineering, Islamic Azad University, Qom Branch, Qom, Iran ahmad sharif Assistant Professor, Faculty of Computer Engineering, Islamic Azad University, Qom Branch, Qom, Iran ali karimi Ph.D. Candidate, Department of Computer Engineering, Islamic Azad University, Qom Branch, Qom, Iran With the swift advancement of the Internet of Things (IoT), Vehicular Ad Hoc Networks (VANETs) have become a crucial component in enabling smart transportation systems by supporting real-time communication between vehicles and roadside units (RSUs). In these networks, vehicles function as mobile nodes that generate and transmit data across the system. A major challenge in VANETs is ensuring the integrity and trustworthiness of shared messages, as any malicious or inaccurate information could severely impact safety and system performance. This research introduces a trust management framework that integrates VANET with blockchain technology and fuzzy logic to improve the reliability of vehicle-to-vehicle communication. When an event is detected, a vehicle instantly broadcasts a corresponding message. RSUs then evaluate the sender’s trust level and verify the message before validation. To minimize communication overhead and avoid duplication, repeated messages are filtered prior to distribution. Unlike conventional trust models that depend on computationally heavy consensus mechanisms such as Proof of Work (PoW), the proposed system adopts a Chord-based distributed architecture. This approach significantly lowers processing times and boosts scalability. The framework utilizes a multi-phase trust evaluation process involving message scoring, dynamic trust calculation, and formation of evaluator groups by RSUs. Simulations reveal notable gains in message credibility: a 6% increase compared to the Score-Based Trust Management System (SBTMS) and an 11% improvement over PoW-based approaches. These results underline the effectiveness of the proposed model in achieving a balance between security, scalability, and low latency in VANET environments. By merging VANET architecture with decentralized trust mechanisms and soft computing techniques, this study presents an innovative and pragmatic solution to one of the key challenges in vehicular communications—facilitating secure, efficient, and trustworthy message exchange in highly dynamic, distributed networks.

Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 22 4 2026 3 1 Predicting glioma brain tumor grades using ensemble machine learning 122 101 FA Hojjat Emami Associate Professor, Department of Computer Engineering, Faculty of Engineering, University of Bonab, Bonab, Iran Babak Azarnavid Assistant Professor, Department of Mathematics and computer science, Basic Science Faculty, University of Bonab, Bonab, Iran Mohsen Abdolhosseinzadeh Assistant Professor, Department of Mathematics and computer science, Basic Science Faculty, University of Bonab, Bonab, Iran Gliomas, or in other words, aggressive and progressive brain tumors, lead to great complexity in the diagnosis and treatment of patients. While recent machine learning models provided encouraging results in glioma diagnosis and grading, the topic is open, and more efforts are needed. Existing models, despite encouraging results, often fall short of the ideal diagnostic state, highlighting the need for further research to develop robust and high-performing predictive models . This study introduces an optimized ensemble machine learning (EML) model designed to maximize classification (grading) performance and mitigate the pervasive issue of overfitting in glioma grading. Our approach employs a two-layer architecture that synergistically combines diverse weak and base learners. In the first layer, a diverse set of learners, including support vector machine (SVM), categorical boosting (CatBoost), extremely randomized trees (ERT), and random forest (RF), is integrated. This initial ensemble aims to capture a broad spectrum of grading patterns and enhance the overall accuracy by leveraging the complementary strengths of each base model. The outputs from this first layer, representing diversified classification probabilities, are then fed into a second-layer logistic regression (LR) model. This layer refines the predictions, performing the ultimate classification while explicitly addressing and eliminating the overfitting problem, thereby promoting better generalization to unseen data. To rigorously evaluate the performance of the proposed ELM model, a comprehensive comparison was conducted against its constituent base learners and counterpart machine learning models. All models were assessed using a standard, publicly available glioma dataset. To prevent overfitting, examine the robustness of models, and evaluate models fairly, a 5-fold cross-validation strategy is used in experiments. The effectiveness of models was measured using four performance metrics, including accuracy, recall, precision, and F1-score. The experimental results demonstrate the superior performance of the proposed EML model. Across all evaluated metrics, our model consistently outperformed the individual base learners and other benchmarked algorithms, securing the top rank in terms of accuracy. Specifically, the LR model operating on the first-layer ensemble predictions proved highly effective in both enhancing accuracy and preventing overfitting. Following our proposed model, the standalone LR and RF models demonstrated commendable performance, ranking second and third, respectively . The findings of this study underscore the significant potential of an optimized EML model for advancing the field of glioma tumor grading. The proposed model generated promising results and mitigated overfitting through integrating diverse base learners and using an LR model as a meta-model. The results reveal that the proposed model is a reliable and robust tool that can aid Clinical specialists in effectively diagnosing and classifying gliomas, ultimately paving the way for improved patient satisfaction.