A practical and privacy-preserving framework for real-world large language model services

Wireless bandwidth is in greater demand than ever before due to the Internet of Things' (IoT) applications' rapid expansion in fields including smart cities, autonomous and Industry 4.0. Traditional fixed spectrum allocation approaches can lead to inefficient utilization and excessive interference levels, particularly in densely populated areas. The purpose of this evaluation is to create an intelligent, decentralized, and privacy-preserving framework for optimizing spectrum detection and sharing among IoT devices utilizing machine learning (ML) techniques. The Cognitive Radio Networks (CRNs) Dataset is gathered from the Kaggle source. The procedure consists of four sequential steps. Each IoT node uses Extreme Gradient Support Vector with Adaptive Ant Colony Optimization (EGSV-AACO) to monitor spectrum occupancy and identify idle bands. Each node builds a local spectrum access model based on temporal spectrum patterns. Model weights are delivered to a nearby edge server on a regular basis to avoid exposing raw data using Federated Averaging (FedAvg). The server aggregates the locally trained models to form a global model and redistributes it to all participating devices. This updated global model will drive real-time, collision-free spectrum allocation among IoT devices. A smart campus simulation using MATLAB shows that the proposed EGSV-AACO framework ensures access convergence, improves spectrum usage, and prevents raw data leakage. The developed model outperforms all baseline methods and achieved an accuracy of 97%, precision of 97.5%, recall of 96%, and an F1-score of 96.5%. Overall, this research introduces a novel Federated EGSV-AACO framework that significantly enhances decentralized, privacy-preserving, and intelligent spectrum sensing and sharing in IoT networks.

Abstract Cybercriminals continually advance their tactics, exploit novel attack vectors, and target emerging technologies. To mitigate these risks, cyber insurance policies must remain abreast of the latest technological developments. Staying technologically updated enables an Insurance Company (IC) to assess risks more precisely, tailor policies to a potential potential Policyholder (PH), and accurately calculate fair premiums. By incorporating innovative risk assessment methodologies, robust Know-Your-Customer protocols, and automated claims-handling processes, ICs can offer tailored and cost-effective solutions to their PHs. Embracing technological advancements enables the field of cyber insurance to adapt to the ever-changing landscape of cyber threats, providing comprehensive protection to organizations. This article introduces , an innovative privacy-preserving framework designed to deliver robust security and privacy assurances to PHs against honest but inquisitive entities within the cyber insurance ecosystem. also equips ICs with automated processes for claims management. At the core of lies Hyperledger Aries, leveraging verifiable credentials to empower PHs with identity ownership and data control. Our framework is complemented by Hyperledger Fabric, which imbues with intelligent functionalities enabling PHs and ICs to autonomously execute actions related to cyber insurance and gather historical cybersecurity data. In this direction, we have successfully implemented the key components of and conducted a quantitative performance assessment. We also substantiate its security and privacy attributes, confirming that effectively achieves its objectives. In summary, represents a forward-thinking solution poised to enhance cyber insurance in an era of ever-evolving cyber threats, offering a promising avenue for safeguarding organizations and policyholders in the digital landscape.

A hybrid framework for privacy preservation in next-generation consumer electronics networks

Blockchain-Enabled Privacy-Preserving Framework for Secure Electronic Health Records Sharing and Diagnosis in Internet of Medical Things

The widespread adoption of the Internet of Things (IoT) has revolutionized various sectors, including healthcare and transportation, by facilitating extensive data gathering and the provision of advanced, intelligent services. However, this growth also amplifies the risks of privacy breaches, unauthorized access, and resource exhaustion, particularly in constrained devices that cannot afford heavy cryptographic operations. Existing solutions often compromise between efficiency and security, leaving systems exposed to replay, Man-in-the-Middle, and even quantum-era threats. This paper proposes a novel authentication and privacy-preserving framework tailored for resource-constrained IoT environments. The design integrates multi-phase processes, including registration, key generation, encryption, mutual authentication, verification, and secure data retrieval. The framework leverages physical-layer features such as RSSI and LQI for enhanced authentication accuracy, supported by cryptographic primitives like hashing and elliptic curve operations. Experimental evaluation using a large-scale IoT dataset demonstrates consistent encryption times between 0.01 and 0.10 seconds, stable latency performance, minimal memory consumption of 0.497 MB, and a detection rate of 0.85. Comparative analysis shows superior efficiency over baseline models in terms of computational overhead and resilience. The results confirm that the proposed scheme provides a robust yet lightweight security architecture, paving the way for secure IoT deployments in latency-sensitive and resource-limited applications.

Abstract Federated Learning (FL) offers a privacy-preserving framework suitable for sensitive domains such as healthcare. This study aims to investigate how different types and intensities of anomalous data introduced by individual clients affect overall model performance in cross-silo FL environments. Additionally, it explores whether models analyzing gradient representations can detect such anomalies during training. We conduct systematic experiments injecting six types of anomalies at varying strengths into training data from two distinct datasets. Performance degradation is measured and statistically analyzed. Furthermore, we develop a Variational Autoencoder (VAE) trained on clean gradient representations to detect deviations caused by anomalies. Our findings indicate that the impact of anomalies on model accuracy varies significantly across datasets and anomaly types. CIFAR-10 data shows higher sensitivity compared to the biological cellular data derived from Quantitative Phase Imaging (QPI). The VAE-based gradient anomaly detection successfully identifies subtle shifts in gradient distributions, but effective differentiation is observed primarily for the QPI data. The results emphasize the importance of tailoring FL robustness and anomaly detection strategies to specific datasets and anomaly characteristics. Gradient-based detection methods show promise for enhancing FL security, but require further refinement. This work contributes critical insights for designing more reliable and secure FL systems, particularly in sensitive domains like healthcare.

Real-world person re-identification (Re-ID) systems are susceptible to malicious attacks, leading to the leakage of pedestrian images and the Re-ID model, posing severe threats to the privacy of both system owners and pedestrians. Existing privacy-preserving person re-identification (PPPR) methods fail to simultaneously resist data leakage, model leakage, and data & model leakage while compromising the normal functionality of Re-ID systems. In this paper, we begin with an in-depth analysis of prior methodologies and identify the gap between existing works and the ideal PPPR paradigm. Inspired by the concept of 'Let the invisible perturbation become the system trigger', we propose SHIELD, a pioneering and comprehensive two-stage privacy-preserving framework. To resist data leakage, we propose a self-supervised method for Protected Dataset Generation in the first stage, which obviates the dependence on identity labels and ensures image quality. To resist model leakage without compromising the normal retrieval accuracy, we propose Original Feature Deconstruction and Protected Feature Alignment to train the system model with paired protected and original images. Extensive experiments substantiate that SHIELD significantly outperforms existing PPPR methods, offering robust and holistic protection for Re-ID systems while maintaining decent retrieval accuracy for authorized users. The code will be released soon.

Medical artificial intelligence (AI) systems depend heavily on high-quality data representations to support accurate prediction, diagnosis, and clinical decision-making. However, the availability of large, well-annotated medical datasets is often constrained by cost, privacy concerns, and the need for expert labeling, motivating growing interest in self-supervised representation learning. Among these approaches, contrastive learning has emerged as one of the most influential paradigms, driving major advances in representation learning across computer vision and natural language processing. This paper presents a comprehensive review of contrastive learning in medical AI, highlighting its theoretical foundations, methodological developments, and practical applications in medical imaging, electronic health records, physiological signal analysis, and genomics. Furthermore, we identify recurring challenges, including pair construction, sensitivity to data augmentations, and inconsistencies in evaluation protocols, while discussing emerging trends such as multimodal alignment, federated learning, and privacy-preserving frameworks. Through a synthesis of current developments and open research directions, this review provides insights to advance data-efficient, reliable, and generalizable medical AI systems.

Federated learning (FL) is a disseminated machine learning (ML) paradigm that gained significant consideration in modern days, particularly in a domain of the internet of things (IoT). FL saves communication bandwidth when compared to centralized ML processes by eliminating the need to transmit raw client data to a central server, thereby enhancing data privacy. Nevertheless, participant privacy is still compromised through inference attacks and similar threats. Additionally, a data excellence provided through clients can differs significantly, and excessive inclusion of low-quality data during training may degrade the overall performance of the global model. Hence, this research introduces a gradient descent optimization assisted weighted federated learning (GDO-WFL) method for privacy preservation. The proposed GDO-WFL approach is significantly efficient as it strengthens privacy preservation through reducing exposure to inference attacks and optimises gradient updates for secure learning. Through weighting client contributions based on data quality, an undesirable effect of low-quality data can be minimised, helping to maintain a strength as well as accuracy of the global model. The experimental results illustrate a proposed GDO-WFL approach maintains an overall accuracy of 99.3 and 91.5% on MNIST and CIFAR-10 datasets as compared to the existing method of FedlabX method.

The rapid growth of cloud computing and big data analytics has intensified concerns over privacy when sensitive data are outsourced to third-party cloud providers. Traditional encryption techniques protect data confidentiality but significantly limit the ability to perform expressive and efficient queries, particularly in distributed and multi-cloud environments. Motivated by the increasing demand for secure analytics across healthcare, finance, IoT, and collaborative cloud platforms, this review systematically examines privacy-preserving query processing techniques for encrypted data in multi-cloud settings. Following PRISMA guidelines, a systematic literature review of published peer-reviewed studies is conducted. The reviewed approaches are categorized into homomorphic encryption-based methods, searchable encryption techniques, secure multi-party computation, trusted execution environments, and hybrid architectures. The analysis highlights key trade-offs among privacy guarantees, query expressiveness, computational efficiency, and scalability. While hybrid and multi-cloud approaches improve flexibility and fault tolerance, they introduce new challenges related to leakage, communication overhead, and trust assumptions. This review identifies critical research gaps, including limited real-time support, side-channel vulnerabilities, and the absence of standardized benchmarks. Finally, future research directions are outlined, emphasizing AI-assisted encrypted querying, federated analytics, and post-quantum privacy-preserving frameworks for multi-cloud environments.

Mapping single-cell datasets to large atlases is often hindered by server constraints and privacy concerns. We present CytoVerse, a framework that runs scRNA-seq Foundation Models (scFM) entirely in the browser. Three key contributions enable this: (1) deploying models via ONNX without server side compute; (2) using compressed indexing (IVFPQ) to search a more then 20 million cell reference from the client; and (3) a lightweight protocol for sharing embeddings across consortia without exposing raw data. CytoVerse thereby provides a scalable, privacy preserving framework for distributed single-cell analysis.

Driver drowsiness is a serious concern for road safety within intelligent transportation systems, and it can undermine the safety and dependability of critical transport infrastructure. As modern vehicles become more connected and data-focused, centralized learning systems that share driver and vehicle information can expose private details and raise privacy and security concerns. This study presents a privacy-preserving framework that enables secure learning among multiple vehicles without sharing raw data. It uses the On-Board Diagnostic-II sensor data, combined with transfer learning, to detect driver drowsiness in real time within a federated learning framework. Signals such as speed, engine revolutions, throttle position, and steering torque are extracted from cars and then converted into image representations using Mel-Frequency Cepstral Coefficients so the model can identify changes in driving behavior. These image features are used to train a pretrained ResNet50 network; this trained model can classify driver states as drowsy or normal. Each vehicle trains on its own data while the central server updates the shared model weights through a client-weighted averaging strategy that keeps learning balanced for all clients. This process keeps data private while the model trained on different driving pattern. Using client weights DrowsyXnet achieved 98.29% accuracy, which is nearly matched the centralized baseline of 98.67%. The latent feature graph showed a clear separation between drowsy and normal states, indicating that the model learns the underlying signals rather than merely incidental correlations. The proposed framework improves intelligent transportation systems while preventing leakage of private data. The use of driver drowsiness detection system into vehicles can prevent drowsiness related accidents and enhance overall road safety.

Lightweight Laplacian Steganography: A Mathematical Framework for Edge-based Privacy Preservation

Assessing data quality in federated health data systems presents unique challenges, particularly when data custodians cannot expose raw data due to privacy regulations. Traditional quality assessment approaches often require centralised access, which conflicts with the principles of data sovereignty and confidentiality. In this study, we evaluate the utility of federated data quality assessment with differential privacy techniques to safeguard sensitive health data. The aim is to develop tooling and demonstrate a proof-of-concept implementation over a synthetic dataset of observational medical data. We present a privacy-preserving framework for evaluating data quality in federated environments using differential privacy. Our approach enables individual data providers to compute local quality metrics and share only aggregated, privacy-protected results. We implement a proof-of-concept that supports predefined quality checks across different data models and demonstrate how meaningful insights into data quality can be obtained without compromising sensitive information. This work demonstrates that differential privacy can be effectively applied to enable federated quality assessment in health data networks without compromising individual privacy. By implementing a proof-of-concept system over synthetic health data, we show that it is possible to obtain meaningful quality metrics in a decentralised setting.

Purpose This study aims to address critical security gaps in distributed systems by developing and evaluating a novel framework that integrates Federated Transfer Learning (FTL), Generative AI (GenAI) and Blockchain technology. Research specifically enhances the detection and mitigation of sophisticated network attacks, including distributed denial of service, man-in-the-middle and model poisoning. Design/methodology/approach The study uses a quantitative evaluation-based approach. The authors propose a three-tiered (Cloud, Edge, IoT) architecture in which edge devices collaboratively train a global intrusion detection model using FTL, ensuring data privacy. GenAI synthesises novel attack data, significantly improving the model’s capability to detect zero-day threats that conventional methods may miss. Blockchain technology secures the integrity of the federated learning process, using a reputation-based mechanism to safeguard against malicious contributions and model poisoning attacks. The framework’s performance is rigorously validated using four public NetFlow-based data sets. NF-ToN-IoT-v2, NF-CSECIC-IDS2018-v2, NF-UNSW-NB15-v2 and NF-BoT-IoT-v2. Findings The experimental results demonstrate the high efficacy of the framework. It achieved a detection accuracy of up to 92% and an F1-score exceeding 80% on client nodes, showing robust performance across heterogeneous and non-IID data distributions (Zhao, 2020; Nguyen et al., 2021). The federated model exhibited stable convergence over 20 aggregation rounds, confirming its adaptability and ability to generalise effectively across diverse domains without centralising sensitive data. The integration of GenAI and Blockchain substantially enhanced the model’s robustness, adaptability and trustworthiness. Originality/value This research presents a novel, holistic security solution synergistically combining FTL, GenAI and Blockchain. Its originality lies in its integrated architecture that simultaneously addresses data privacy, model integrity and adaptability to evolving threats. The findings offer a practical blueprint for creating scalable, privacy-conscious and resilient security frameworks applicable to complex distributed environments.

Fragmented and locally siloed data limit progress in critical care research and education. The European Health Data Space (EHDS) proposes a federated, privacy-preserving framework to connect intensive care units (ICUs) across Europe. Sepsis is an ideal model condition given its heterogeneity, high mortality, and persistent gaps in standardization and outcomes. This narrative review explores how federated and synthetic data can transform sepsis research, quality improvement, and education within the EHDS. It aims to outline both the opportunities and practical limitations of building a European-wide, learning ICU network. Recent literature, European policy documents, and federated data initiatives were reviewed to synthesize conceptual, technical, and ethical aspects of implementing federated learning in intensive care. Federated infrastructures enable joint analysis of distributed ICU data without sharing patient-level information, supporting benchmarking and surveillance while maintaining privacy. Synthetic data add value for simulation, algorithm testing, and training but cannot replace real-world complexity. Major barriers include data harmonization, interoperability, and governance. Ongoing projects demonstrate that transparent, secure frameworks can make responsible data sharing feasible. The EHDS offers a realistic foundation for connecting ICUs across Europe through ethically governed federated systems. Combining clinical, engineering, and data science expertise will be key to transforming fragmented ICU information into shared intelligence that supports sepsis research, education, and personalized critical care.

Smart mobility services generate large volumes of sensitive location and identity data, raising critical concerns related to privacy leakage, ‎security vulnerabilities, and trust in large-scale urban deployments. To address these challenges, this paper proposes a blockchain-based ‎privacy-preserving framework for smart mobility services that integrates geo-indistinguishability, pseudonymous authentication, Zero-‎Knowledge Proofs (ZKPs), and Proof-of-Authority (PoA) consensus into a unified architecture. The framework ensures end-to-end privacy ‎by combining calibrated location obfuscation with decentralized transaction validation and immutable auditability, thereby mitigating both ‎inference-based attacks and reliance on centralized trust.‎ The proposed framework was evaluated using the TAPAS Cologne mobility dataset, comprising 1,000 simulated vehicles and 20 block-‎chain validators. Experimental results demonstrate that adversarial inference accuracy is reduced to below 12%, while approximately 75% ‎navigation utility is preserved at balanced privacy budgets. Security analysis confirms robust protection against tracking, replay, Sybil, and ‎collusion attacks, with replay attack success rates reduced from 70% to 2% through the enforcement of timestamps and nonces, along with ‎cryptographic verification.‎ Performance evaluation demonstrates that the framework achieves high throughput (1,200 transactions per second) with sub-second latency ‎‎(0.8 seconds) under realistic transaction loads. Storage growth is optimized to 2.1 GB per million transactions, and the PoA consensus ‎mechanism achieves approximately 30% lower energy consumption compared to Proof-of-Stake-based designs. In addition, resilience ex-‎periments confirm Byzantine fault tolerance under up to 30% malicious validator participation, without service degradation.‎ Overall, the results demonstrate the practical feasibility of deploying the proposed framework in real-world smart mobility ecosystems that ‎require simultaneous privacy preservation, scalability, and energy efficiency. The framework represents a significant step toward trustwor-‎thy, privacy-aware, and sustainable smart-city mobility infrastructure, providing a robust foundation for next-generation decentralized mo-‎bility services‎.

BackgroundDiabetes prediction requires accurate, privacy-preserving, and scalable solutions. Traditional machine learning models rely on centralized data, posing risks to data privacy and regulatory compliance. Moreover, health care settings are highly heterogeneous, with diverse participants, hospitals, clinics, and wearables, producing nonindependent and identically distributed data and operating under varied computational constraints. Learning in isolation at individual institutions limits model generalizability and effectiveness. Collaborative federated learning (FL) enables institutions to jointly train models without sharing raw data, but current approaches often struggle with heterogeneity, security threats, and system coordination.ObjectiveThis study aims to develop a secure, scalable, and privacy-preserving framework for diabetes prediction by integrating FL with ensemble modeling, blockchain-based access control, and knowledge distillation. The framework is designed to handle data heterogeneity, nonindependent and identically distributed distributions, and varying computational capacities across diverse health care participants while simultaneously enhancing data privacy, security, and trust.MethodsWe propose a federated ensemble learning framework, FedEnTrust, that enables decentralized health care participants to collaboratively train models without sharing raw data. Each participant shares soft label outputs, which are distilled and aggregated through adaptive weighted voting to form a global consensus. The framework supports heterogeneous participants by assigning model architectures based on local computational capacity. To ensure secure and transparent coordination, a blockchain-enabled smart contract governs participant registration, role assignment, and model submission with strict role-based access control. We evaluated the system on the PIMA Indians Diabetes Dataset, measuring prediction accuracy, communication efficiency, and blockchain performance.ResultsThe FedEnTrust framework achieved 84.2% accuracy, with precision, recall, and F1-score of 84.6%, 88.6%, and 86.4%, respectively, outperforming existing decentralized models and nearing centralized deep learning benchmarks. The blockchain-based smart contract ensured 100% success for authorized transactions and rejected all unauthorized attempts, including malicious submissions. The average blockchain latency was 210 milliseconds, with a gas cost of ~107,940 units, enabling secure, real-time interaction. Throughout, patient privacy was preserved by exchanging only model metadata, not raw data.ConclusionsFedEnTrust offers a deployable, privacy-preserving solution for decentralized health care prediction by integrating FL, ensemble modeling, blockchain-based access control, and knowledge distillation. It balances accuracy, scalability, and ethical data use while enhancing security and trust. This work demonstrates that secure federated ensemble systems can serve as practical alternatives to centralized artificial intelligence models in real-world health care applications.

Federated machine learning is a solution to the problem of identifying instances of credit card fraud in any bank without violating privacy laws. The framework proposed in this paper any bank without violating privacy laws. The framework proposed in this paper introduces a privacy-preserving model that combines Federated Machine Learning (FML) with cutting-edge techniques such as differential privacy, homomorphic encryption, and secure multi-party computation (SMPC). The framework, which uses simulated financial transaction data sourced from publicly accessible datasets (IEEE-CIS and PaySim), trains local models at each financial organization and aggregates them into one global model. The system proposed can detect fraud 5–10% better than traditional single-bank models while at the same time providing high privacy standards. The usage of SMPC and homomorphic encryption avoids sensitive data sharing while differential privacy secures the system against data leakage in case of an attack. The nature of the results of the proposed system have encouraged regulatory authorities to pursue the route of the federated learning framework to detect fraud while simultaneously remaining on the right side of privacy laws like GDPR and DPDP. This framework is suitable for the banking sector and other industries that encounter similar privacy and security problems because of its adaptability to changing data protection legislation. The integration of multiple privacy-preserving technologies into a federated system for financial fraud detection is a significant contribution of this research.