Authors - Shreepreet Sahu, Prasant Kumar Sahu Abstract - Free-space optical (FSO) communication is a promising technology for B5G and 6G communication systems due to its security, reliability, high data rates, low latency and electromagnetic immunity. However, its performance is limited by atmospheric turbulence, weather conditions, beam divergence, misalignment errors and link range variations. Existing analytical or simulation-based methods become too complex or computationally expansive as number of impairments considered simultaneously increases introducing a gap in fast and precise system-level performance estimation. This limitation motivates the use of intelligent data-driven approaches capable of capturing highly nonlinear interrelations. This paper proposes an artificial neural network (ANN) for predicting Q-factor values of the modelled system. The ANN-based model is trained by an extensive dataset generated under varying FSO link ranges and other scenarios. Model legitimacy specification starts with error histograms proceeding through mean squared error (MSE) convergence finding concluding regression analysis before eye pattern evaluation takes place. As shown by the results the high prediction accuracy, generalization capability and closeness of forecasted Q-value to the actual one ensures noticeable improvement over existing framework satisfactorily addressing the above issues. The proposed approach provides an efficient alternative to conventional analytical methods, making it suitable for real-time performance evaluation and optimization practical FSO systems.
Authors - Poorna Pragnya H, Neha V Malage, Pranav Muppuru, Sanya Vashist, Surabhi Narayan Abstract - This work introduces a novel Sequence-to-Sequence (Seq2Seq) framework that converts Electroencephalography (EEG) signals and related metadata into coherent natural language descriptions. The key innovation is a spatio-temporal EEG encoder built using Dense Graph Convolutional Networks (GCNs), which effectively model spatial relationships among electrodes as well as their temporal dynamics in multi-channel EEG data. This encoder is coupled with an attentiondriven Gated Recurrent Unit (GRU) decoder to generate textual sequences. To strengthen learning, the model adopts a multi-task objective that simultaneously predicts scene-level attributes, such as colors and objects, alongside caption generation, promoting better alignment between EEG features and language outputs. Experiments on a large-scale dataset demonstrate competitive results, achieving a BLEU score of 0.21, ROUGE-1 of 0.4519, and ROUGE-L of 0.4447. The generated captions are further used as inputs to a text-to-video generation module. While precise pixel-level matching remains difficult, evaluation shows strong semantic alignment between generated and reference videos, with an SSIM of 0.19 and a CLIP-based semantic similarity score of 0.746. Overall, the results highlight the promise of GCN-based EEG representations for complex language decoding and downstream video generation tasks.
Authors - Barsa Priyadarshani Behera, Monalisa Jena, Ranjan Kumar Behera, Sung-Bae Cho Abstract - Drought prediction remains challenging due to complex physical interactions and limited observability of land-atmosphere processes. This study proposes a Causal-Chain Transformer that explicitly employs drought evolution through three sequential latent representations corresponding to heat stress, evaporation stress, and soil moisture stress. Using only past temperature and evaporation data over a xed historical window, the model predicts future drought occurrence at a predened lead time, while excluding current soil moisture to avoid target leak- age. Experiments on region-averaged NASA POWER and ERA5-Land datasets over Odisha, a state of India, show that the proposed model achieves the highest F1-scores (0.709 on NASA POWER and 0.467 on ERA5-Land), outperforming logistic regression, Long Short-Term Memory (LSTM), and standard Transformer baselines. The learned latent stress signals provide intrinsic interpretability, with early increases in heat and evaporation stress frequently preceding observed drought events, supporting its applicability for early-warning systems in agriculture- dependent regions.
Authors - Aman Kamboj, Vishal Kumar, Abhishek Mishra, Kalpesh Patil Abstract - Accurate 3D reconstruction of tire tread geometry is essential for industrial applications such as automated wear estimation, defect detection, and quality assurance. However, 3D reconstruction of tires from multi-view RGB images remains challenging due to low-texture rubber surfaces, repetitive groove patterns, and sensitivity to lighting variations. These factors often lead to incomplete or noisy reconstructions when using standard photogrammetry. This paper presents an optimized multi-view tire reconstruction framework tailored specifically for tire tread surfaces. The resulting 3D tire model was compared with the reference 3D model obtained from a laser scanner. The comparison showed a mean point-to-point distance of 0.05mm between the two models, indicating a high level of geometric accuracy and close agreement with the ground-truth laser-scanned model. Experimental evaluations further demonstrate that the our optimized method is fast and achieves higher completeness, depth information, better preservation of tread grooves. Overall, the proposed framework provides an accurate tire 3D reconstruction solution capable of delivering the precision required for modern tire inspection and analysis.
Authors - D. P. Jayathung, M. Ramashini, Juliana Zaini, R. Muller, Liyanage C. De Silva Abstract - The primary objective of this research is to explore and interpret the complex flight kinematics of bats in order to deepen aerodynamic understanding and inspire future technological innovation. To achieve this, the study adopts a hybrid approach for estimating flapping pose phases in high-speed bat flight recordings. Accurately distinguishing between the upstroke and downstroke phases is essential for examining the subtle dynamics and movement patterns of bats’ uniquely flexible wing structures. The methodology followed a structured work-flow, beginning with video acquisition using an array of 50 high-speed cameras that recorded bat flights at 1000 frames per second within a controlled tunnel environment. An enhanced YOLOv5L model was then employed to remove un-necessary frames, achieving a mean Average Precision (mAP) of 99.3% and successfully filtering out more than 85% of unwanted footage. For the pose estimation, this work used DeepLabCut to define 20 anatomical keypoints. After com-paring five backbone architectures, this study selected ResNet50 as the most suit-able model, as it yielded the lowest test RMSE (3.98) and the highest test mAP (97.62%). A rule-based geometric method was developed to classify bat wing-beat phases using elbow–wrist–wingtip angles derived from DeepLabCut key-points. By analyzing the smoothed angle trajectory and its temporal derivative, the rule-based approach reliably identified upstroke and downstroke cycles, which were validated using test videos. The extracted phase information supports a deeper biomechanical understanding of bat flight while also enabling applications in bio-inspired robotics, real-time flight monitoring, and automated analysis of complex animal motion.
Authors - Frances Ysabelle D. Rebollido, Jaime D.L. Caro Abstract - The rapid development of artificial intelligence (AI) creates new opportunities and challenges in introductory programming education. Existing AI tools provide immediate support and feedback to students, but they have the tendency to generate inaccurate, biased, or pedagogically unsuitable responses. To address this, we introduce the Agentic Learning & Adaptation System (ALAS), an Agentic AI-based system designed to deliver tailored and educationally grounded support for students. Hence, with this process, ALAS generates responses that are adaptive, and pedagogically appropriate. This enables ALAS to provide personalized support to students. Its modular design provides a scalable foundation for integrating additional agents and functions. We present the conceptual design and early-stage prototype of ALAS to demonstrate its potential in enhancing students’ learning experiences and supporting the responsible use of AI in computing education. Future work will focus on implementing and evaluating ALAS in a classroom setting.
Authors - Yazhiniyan Tamizhnambi, Senthil Prakash P.N Abstract - Having trustworthy systems in a decentralized systems remains a challenge, especially in adversarial conditions that include model poisoning, sigil attacks and unauthorized re-entries. Despite the fact that federated learning and swarm learning can achieve collaborative model training without sharing raw data, existing methodologies largely use fixed identities, self-reported accuracy, or direct weight comparison, which in an open or semi-trust environment is likely to be weak. This work presents a blockchain-based trust system in swarm learning, based on behavioral fingerprinting instead of identity-based accountability. In the suggested system, all involved nodes produce a behavioral fingerprint at every training round, which contains an accuracy of the challenge-sets, deviation of updating the model, and the distribution of features importance. The fingerprints are then stored on chain with the help of Merkle root structures, ensuring transparent behavioral tracking across rounds. To address early-time poisoning and delayed attacks, the system will utilize trust-weighted round-gated aggregation where the model updates will be verified before affecting other participants. Trust is measured through short-term and long-term consistency of behavior supported by Round Performance Score (RPS) which measures inconsistency with peer consensus during a round. The framework further resists Sybil and reentry attacks by matching behavioral fingerprints across identities, ensuring that malicious models cannot bypass detection by resetting node credentials. Behavioral fingerprints are matched across identities to stop further Sybil and re-entry attacks. This ensures credential resetting by nodes to bypass detection, since the behavior of the model will more or less be the same. The experimental analysis of heterogeneous hospital data sets shows improved universal accuracy, adequate poisoned updates mitigation, and dependable detection of malicious re-entry strategies.
Authors - Shipra Swati, Sunita Kumari, Santwana Sneha Abstract - The significant changes in brain dynamics caused by alcohol addiction can be captured by electroencephalography (EEG). Automated alcoholism detection using EEG has gained attention as a non-invasive, objective replace traditional clinical assessments. This study provides a detailed comparison between conventional machine learning models and deep learning architectures for the EEG-based classification of alcoholism. It uses a publicly available multichannel EEG dataset containing recordings of both control and alcoholic subjects. Preprocessing and feature extraction in the time, frequency, and time-frequency domains are done before the assessment of traditional classifiers like k-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM). Furthermore, image-like EEG representations were used to adapt deep convolutional neural networks (ResNet and GoogleLeNet) for classification. According to experimental results, KNN achieves competitive accuracy with little training time, while ensemble methods and deep residual networks perform better than simpler classifiers. The results demonstrate the relative benefits and drawbacks of deep learning and statistical learning paradigms for EEG-based alcoholism detection.
Authors - Raghav, Chanchal Maurya, Sunakshi Singh Abstract - Smart home ecosystems consist of resource-constrained IoT devices that continuously generate sensitive data, making privacy protection, access control, and resilience to device compromise critical challenges. This paper proposes a privacy-preserving one-way access-control encryption scheme for cloud-assisted smart home environments, designed to enforce a strict separation between data generation and data access. In the proposed scheme, devices are granted encryption capability only, while decryption authority remains exclusively with the device owner, thereby preventing unauthorized data disclosure and eliminating key escrow risks. To protect identity privacy, devices employ periodically refreshed pseudonymous identifiers derived from ephemeral secrets, ensuring unlinkability and resistance to tracking and profiling attacks. The scheme further limits the impact of device compromise and prevents adversarial data injection. Performance evaluation demonstrates that the proposed scheme incurs lower computational and communication overhead than existing encryption schemes, making it lightweight and well suited for resource-constrained smart home IoT deployments.
Authors - Anudeep Arora, Minal Maheshwari, Abha Pandey, Neha Chabra, Prashant Vats, Surbhi Sharma Abstract - The rapid expansion of e-commerce platforms has intensified exposure to sophisticated digital threats, including deepfake-driven identity manipulation, financial fraud, and large-scale automated attacks that undermine consumer trust. Traditional rule-based and signature-driven security mechanisms are increasingly inadequate against adaptive and AI-generated adversarial behaviors. This paper investigates the role of artificial intelligence in enabling proactive threat detection and sustained trust preservation within modern e-commerce ecosystems. We present an AI-enabled security framework that integrates deep learning-based anomaly detection, behavioral analytics, and multimodal content verifi cation to identify fraudulent transactions, synthetic media attacks, and coordinated threat patterns in real time. The proposed approach leverages temporal user behavior modeling, transaction graph analysis, and fea ture-level risk aggregation to enhance detection accuracy while minimiz ing false positives. Additionally, explainable AI components are incor porated to support transparency and regulatory compliance, thereby re inforcing user confidence and platform accountability.
Authors - Atharva Patil, Dibyanshu Singh, Tanish Dadarkar, Suman Madan Abstract - The use of artificial intelligence in the automotive system presents legal, ethical, and societal issues such as accountability, safety, human trust, and data privacy. In the case of system failure, explainable behaviour, necessitating the complexity and opacity of AI-driven decision-making. Bias in the training dataset may cause unequal system performance in different traffic environments and road uses, thus the need for representative data and validation. The vast amount of vehicle and data collected raises privacy issues, thus the need for secure data handling and anonymization. Ethical system design should therefore consider fairness, safety, and accountability as primary engineering constraints for responsible AI-enabled vehicle deployment. They deliver safe, more efficient and sustainable vehicles and services. Not only are the vehicles themselves being modernized through the technology, but manufacturing processes and supply chain management on the backend are also changing.
Authors - Sapna Jyoti Patel, Sumit Kumar Debnath Abstract - This paper analyses the blindness property of two code-based blind signature schemes: one by Chen et al. [17] and the other one by Ren et al. [19]. Both [17] and [19] claimed that their protocols provide blindness under brute force attacks. Through detailed analysis, this paper demonstrates that the aforementioned code-based blind signature schemes (CBBSS), in practice, do not satisfy the property of blindness. Moreover, we use a zero-knowledge proof of knowledge (ZKPK) in [17] and [19] in order to achieve the blindness property.
Authors - Haitham Al Habsi, Norliza Mohamed, Suriani Mohd Sam, Hazilah Mad Kaidi, Norulhusna Ahmad Abstract - RFID-enabled IoT systems have transformed supply chain traceability, yet their authentication mechanisms remain critically exposed. Common threats include tag cloning, replay attacks, rogue reader exploitation, and centralized database breaches. This paper examines authentication weaknesses through a five-layer IoT architectural model, identifying four root causes: weak encryption, static identifiers, absent mutual authentication, and over-reliance on centralized trust. These weaknesses are mapped across physical, connectivity, middleware, analytics, and application layers to illustrate how failures propagate systemically rather than in isolation. In response, a blockchain assisted authentication framework is proposed, combining lightweight cryptographic primitives, immutable audit logging, and smart contract-driven access control to eliminate single points of failure. Comparative analysis confirms that decentralized architectures substantially reduce replay and cloning risks while remaining compatible with existing RFID infrastructure. The findings offer a practical analytical foundation for building resilient, adaptive authentication in next-generation IoT anti-counterfeit deployments.
Authors - Timothy T Adeliyi, Debajit Saikia Abstract - Banks rely heavily on long-term customer relationships to ensure sus-tainability, profitability, and competitive advantage. In an increasingly saturated financial services market, customer churn poses a significant threat to revenue stability. Artificial intelligence (AI) and machine learning (ML) have enhanced predictive capabilities in churn modelling; however, the increasing complexity of high-performing models often limits human interpretability and trust. This study investigates how predictive accuracy can be balanced with interpretability in credit card churn modelling through an explainable machine learning frame-work. A quantitative mono-method design was adopted using a publicly available credit card churn dataset comprising approximately 10,000 customer records. Following exploratory data analysis (EDA), multiple classification algorithms were implemented, including logistic regression, decision trees, k-nearest neigh-bours, support vector machines, gradient boosting, and random forests. The ran-dom forest model achieved the highest predictive performance (AUC = 0.940753) and was subsequently selected for interpretability analysis using Shap-ley Additive exPlanations (SHAP). The SHAP-based analysis enabled transpar-ent identification of feature importance and revealed the underlying drivers in-fluencing churn predictions. Graphical explanations were generated to enhance human understanding and support decision-making processes. The findings demonstrate that sustainable deployment of ML systems in banking requires a deliberate integration of predictive performance, domain knowledge, human-in-the-loop validation, and continuous monitoring. This study contributes to the dis-course on trustworthy AI in financial analytics by illustrating how interpretability techniques can strengthen confidence in high-performing churn prediction mod-els without compromising accuracy.
