Next-gen and autonomous federated systems K. S. Divya, K. Soumya Federated Learning in Finance Unlocking Privacy Preserving and Cyber Resilience Using AI, 2026 Federated learning (FL) has developed over the last ten years into an innovative approach that facilitates collaborative model creation while maintaining data privacy protection. Although traditional federated systems work well, they have significant difficulties conducting large-scale operations, adjusting to new surroundings, and making judgments in complicated financial circumstances without human intervention. The financial services sector must create new federated systems because it requires intelligent, safe, real-time data analysis for networks that are decentralized. The chapter examines autonomous federated systems by outlining their technological elements, key advances, and applications in the financial industry. The research proposes a future direction that builds self-optimizing federated networks that preserve operational effectiveness and trust by fusing AI coordinating systems with quantum-resistant security mechanisms and continuous learning systems.
Graph Neural Networks and Hybrid Deep Learning for Personalized Anomaly Detection and Health Risk Forecasting in IoT Healthcare Infrastructures Poongothai P, Parvathi P, Soumya K, Divya K S, Prathap G, Narikamalli Yaswanth 2025 International Conference on Computing and Communications Computingcon 2025, 2025 Continuous monitoring of multimodal biosignals on IoT edge devices can enable early detection of physiological deterioration, yet existing methods struggle to capture the nonlinear, non-contiguous relationships that characterise real-world vital signs and must still meet strict latency and power budgets. We introduce a lightweight hybrid architecture that turns each <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5 0}$</tex>-sample window of electromyography, galvanic-skin response, heart-rate, and blood-pressure data into a k-nearest-neighbour graph, applies two graph-convolutional layers to encode structural affinities, and refines the resulting node features with a bidirectional LSTM. A dual-head decoder simultaneously (i) reconstructs the window-yielding an anomaly score via reconstruction error-and (ii) forecasts the next-window systolicpressure risk. Trained on a physiologically realistic fifty-subject dataset with 5% injected anomalies, the model attains 97% overall accuracy, a 0.91 ROC-AUC, and a 6.8 mm Hg root-mean-square error in systolic-pressure prediction. Crucially, the entire pipeline executes in 8 ms per window on a Raspberry Pi 4 while consuming under 2 GB of memory, matching cloudscale accuracy at IoT-class cost. These results demonstrate that fusing graph structure with bidirectional temporal context yields clinically meaningful precision within the constrained compute envelopes of edge healthcare devices. Future work will target recall improvements via rare-event augmentation, adaptive thresholds, and validation on large-scale ICU telemetry.
Machine Learning-Based Recommendation Engine for Intelligent Cross-Selling and Upselling in E-Commerce Systems Irfan Abdul Karim Shaikh, L.Vijayakumar, Kishore Bhattacharjee, Rohit Kumar Abhimalla, Vivekanand Pandey, Divya K S 2025 5th International Conference on Emerging Research in Electronics Computer Science and Technology Icerect 2025, 2025 These days, no e-commerce platform is complete without cross-selling and up selling tactics, which boost revenue and delight customers alike. The success of these initiatives depends on accurately gauging consumer opinion. With order to aid with up selling and cross-selling, this study devised a sentiment analysis method for evaluating user-generated material. A comprehensive data cleansing procedure was executed, which comprised tokenization, converting text to lowercase, and removing stop words and punctuation. After the data was cleaned up, it was categorized using the LGCU-CS-DBN model, a DBN based deep learning technique. The GCU-CSO technique was used to optimize the DBN architecture in order to further boost the model’s performance. With this method, they were able to fine-tune the activation functions and hidden-neuron count. Achieving a precision of 94.80%, the optimized model demonstrated high sentiment categorization capabilities. Standard performance metrics-based comparative evaluations validated its superiority over more conventional algorithms. These findings show that the LGCU-CS-DBN model is a powerful instrument for e-commerce systems to employ intelligent up selling and cross-selling methods. Because of its flexibility and high accuracy, the model may be used in real-world applications, allowing businesses to make better recommendations and engage customers more effectively.
Optimizing Lightweight Cryptographic Protocols for Secure Communication in Internet of Things(IoT) Devices Divya K.S, Margaret Mary T, Sreeparna Chakrabarti, Prakash V.S, K. Chitra Proceedings IEEE 5th International Conference on Communication Computing and Industry 6 0 2024 C2i6 2024, 2024 The Internet of Things (IoT) has significantly reshaped technological advancements, integrating a wide range of devices into a unified ecosystem. This interconnected network extends across various sectors, from everyday household gadgets to sophisticated industrial equipment, enabling seamless data exchange and automation. However, the exponential growth of IoT devices introduces critical vulnerabilities, as many devices are inherently resource-constrained and unable to support conventional security mechanisms effectively. This paper presents a novel lightweight cryptographic protocol designed for secure communication in resource- constrained IoT devices. The protocol combines Elliptic Curve Cryptography (ECC) with an optimized symmetric encryption scheme to ensure data confidentiality, integrity, and authenticity while minimizing computational overhead. The proposed protocol is evaluated against common cryptographic attacks and analyzed for its efficiency in terms of computational complexity and energy consumption.
Random Selection for Cloud-Based Load-Distributing in Varied P2P Networks using Enhanced Throttled Algorithm V. S. Prakash, Margaret Mary T, B. N. Shivakumar, Jasmine Gnanadurai, K. S. Divya, Vennila J Proceedings IEEE 5th International Conference on Communication Computing and Industry 6 0 2024 C2i6 2024, 2024 A peer-to-peer network connection is essential to most computer users' daily activities. On heterogeneous P2P networks, maintaining cloud-based load balance is a difficult goal to achieve. Several previous research have provided load rearrangement procedures for additional load diversions, and load division on the point where the original peer arrives. The goal of cloud-based load-sharing in P2P structures is to evenly distribute the burden among net nodes, then it is not effective in efficiently removing traffic that is already happening on the P2P network's routes. By removing the traffic flow and consuming the Random Selection Load Balancing (RSLB) method and the Incredible Store Model, the behaviors in this study are characterized. Using an improved throttled method, the Incredible Store Model is examined to broaden the effects of node variety in addition to agitation toward the capacity circulation in P2P systems. Now different from the current load distribution balancing strategy, the presentation of the Random Selection Load Balancing (RSLB) technique by the Incredible Store Model decreases the network traffic flow pathways with little cloud rate of P2P. By combining the burden in the P2P net system, an investigative and observed conclusion delivers a destruction-free system with information calculated. Peer failure probability in addition to traffic control effectiveness is used to estimate the presentation of the random selection load assessment.
Framework of Multiparty Computation for Higher Non-Repudiation in Internet-of-Things (IoT) Divya K.S, Roopashree H.R, Yogeesh A.C International Journal of Computer Networks and Applications, 2023 Multiparty computation is essential in offering a better form of non-repudiation, which is not much explored in past research work.A review of existing non-repudiation-based approaches found various shortcomings that do not offer a good balance between robust security and algorithm efficiency.Therefore, the proposed study presents a novel yet simple multiparty computation framework to ensure a higher degree of non-repudiation considering a use-case of a highly distributed and large network, i.e., Internet-of-Things (IoT).The study implements a unique encryption mechanism that uses a transformation strategy to perform encoding while using split key management to retain maximal secrecy and multiparty authentication for enhanced security.The simulation outcome of the study showcases that the proposed scheme offers approximately a 48% reduction in computation overhead, 54% minimization in delay, and 58% faster processing in contrast to frequently reported non-repudiation schemes.
Hybrid ML Algorithms for Learning Disability Forecast in School Going Children Using Python in Machine Learning Techniques Margaret Mary T, V S Prakash, Divya K S, Amalorpavam George 7th International Conference on Electronics Communication and Aerospace Technology Iceca 2023 Proceedings, 2023 A learning disability is a neurological illness that impairs a child's ability to read, speak, and do a variety of other skills. The World Health Organization (WHO) estimates that learning disabilities impact 15% of youngsters [14]. The most important challenge for researchers to perform in order to identify learning disabilities early on is efficient prediction and accurate categorization. Our primary goal in this effort is to use soft computing to create a model for the prediction and categorization of learning disabilities. This study proposes a hybrid approach for enabling classification in order to enhance the performance of prediction and classification. This method incorporated classification's primary five techniques. Random Forest, Logistic Regression, Stochastic Gradient Descent, and K-Fold cross validation. In order to implement the system used python. Results analysis reveals the predict of learning disability in effectively
Non-Repudiation-based Network Security System using Multiparty Computation Divya K. S, Roopashree H. R, Yogeesh A C International Journal of Advanced Computer Science and Applications, 2022 —Security has always been a prominent concern over the network, and various essential requirements are required to cater to an efficient security system. Non-repudiation is a requirement about the non-deniability of services acting as a bridge between seamless relaying of service/data and efficient security implementation. There have been various studies carried out towards strengthening the non-repudiation system. There are certain pitfalls that render inapplicability on dynamic cases of vulnerability. The conventional two-party non-repudiation schemes have been widely explored in the existing literature. But this paper also advocates the adoption of multi-party computation, which has better feasibility toward strengthening a distributed security system. The current work presents a survey on the existing approaches of non-repudiation to investigate its effectiveness in the multi-party system. The prime aim of the proposed work is to analyze the current research progress and draw a research gap as the prominent contribution of the proposed study. The manuscript begins by highlighting the issues concerning multi-party strategies and cryptographic approaches, and the security requirements and standardization are briefly discussed. It then describes the essentials of non-repudiation and examines state-of-the-art mechanisms. Finally, the study summarizes and discusses research gaps identified through the review analysis.
