Fine-tuning large language models on downstream tasks is crucial for realizing their cross-domain potential but often relies on sensitive data, raising privacy concerns. Differential privacy (DP) offers rigorous privacy guarantees and has been widely adopted in fine-tuning; however, naively injecting noise across the high-dimensional parameter space creates perturbations with large norms, degrading performance and destabilizing training. To address this issue, we propose DP-SFT, a two-stage subspace fine-tuning method that substantially reduces noise magnitude while preserving formal DP guarantees. Our intuition is that, during fine-tuning, significant parameter updates lie within a low-dimensional, task-specific subspace, while other directions change minimally. Hence, we only inject DP noise into this subspace to protect privacy without perturbing irrelevant parameters. In phase one, we identify the subspace by analyzing principal gradient directions to capture task-specific update signals. In phase two, we project full gradients onto this subspace, add DP noise, and map the perturbed gradients back to the original parameter space for model updates, markedly lowering noise impact. Experiments on multiple datasets demonstrate that DP-SFT enhances accuracy and stability under rigorous DP constraints, accelerates convergence, and achieves substantial gains over DP fine-tuning baselines.

Federated learning (FL) has been increasingly applied in the Internet of Things (IoT), leveraging its decentralized nature to facilitate collaboration among clients and enable resource-constrained clients to jointly train a globally optimal model based on consensus. However, it is difficult to confirm data authenticity and participant integrity due to the unobservability of local training procedures and the inaccessibility of local training data. As a result, FL is highly susceptible to Byzantine attacks, including data poisoning and model poisoning, which can manipulate the training process and degrade model performance. Moreover, IoT data is often highly heterogeneous and high-dimensional, rendering most existing Byzantine-robust FL approaches ineffective in practical scenarios. To address this challenge, we propose GeoFed, which iteratively filters out malicious clients based on the geodesic distance between clients. This geodesic distance is measured on the Riemannian manifold spanned by the covariance of local gradient update. To further mitigate the impact of data heterogeneity, GeoFed assigns a weight factor to each client after removing Byzantine attackers, optimizing the accuracy and flexibility of global model aggregation according to the quality of client data. We conduct extensive experimental evaluations of GeoFed under various Byzantine attack scenarios and highly heterogeneous data environments. To validate the efficacy of GeoFed, we provide a theoretical analysis of its convergence properties. The results demonstrate that GeoFed outperforms state-of-the-art Byzantine-robust FL approaches in heterogeneous IoT settings. Especially, under different Byzantine attacks, the accuracy of detecting malicious clients on the heterogeneous MNIST dataset approaches 100%.

Federated Learning (FL), a distributed machine learning (ML) framework, is susceptible to Byzantine attacks since the attacker can manipulate clients local data or models to compromise the performance of the global model. There has been a wealth of defenses developed to mitigate the attacks by limiting the impact of malicious models. Nevertheless, the attacker can easily circumvent these approaches that rely solely on a single server-side defense, stemming from the high dimensionality of models and the variety of Byzantine attacks. Therefore, we propose Basalt, a Byzantine-robust federated learning framework with a server-client joint defense mechanism that enables multiple clients to train a global ML model under Byzantine attacks. On the client side, we design an efficient self-defense approach with model-level penalty loss that restricts local-benign divergence and decreases local-malicious correlation to prevent misclassification. On the server side, we present an efficient defense strategy based on the manifold and maximum clique, further strengthening the FLs resilience against Byzantine attacks. We provide theoretical guarantees for global model convergence in FL with Byzantine attacks. Our extensive experiments demonstrate that Basalt outperforms existing state-of-the-art works. Especially, it achieves nearly 100% accuracy for detecting malicious clients in nonindependent and nonidentically distributed (Non-IID) MNIST datasets under various Byzantine attacks.

Private learning for vertical decision trees (PVDT) is an emerging paradigm that allows multiple parties to execute cooperative training and inference of decision trees on vertically partitioned datasets, without revealing either party%'s data or model. The state-of-the-art PVDT schemes employ the secret-sharing-based secure multi-party computation (MPC) to admit low computational cost and low bandwidth. Nevertheless, existing schemes need many communication rounds for computing concrete protocols in PVDT, like the less-than comparison, division, etc. This property is not suited for large-communication-latency networks such as WAN. In this work, we present a two-party PVDT framework, calledSwan, to enable a secure, accurate, and fast realization of vertical decision trees. At the core of Swan, we design a secure and parallel protocol for$N$-input multiplication with one communication round. This forms the cornerstone for a series of secure and communication-efficient computation protocols specifically tailored to less-than comparison and division. Along the way, we use these optimized protocols to refine the training and inference processes of PVDT, achieving a significant reduction in both communication costs and rounds. Experimental results show Swan provides top-notch accuracy, and achieves a$10.2\times$and$2.8\times$improvement in online training and inference latency over WAN compared to prior art.

Federated learning (FL) in Internet of Things (IoT) applications facilitates the collaborative training of a global model across distributed devices with a server. Despite its potential, the distributed nature and vulnerability of IoT devices render FL susceptible to Byzantine attacks. Existing approaches to counter these attacks are often impractical in real-world IoT scenarios, mainly due to the challenges posed by nonindependent identically distributed (non-IID) data and the high-dimensional model common in IoT devices. To address these challenges, we propose Guard-FL, an efficient and robust aggregation mechanism assisted by uniform manifold approximation and projection (UMAP) for FL. Guard-FL is designed to enhance the performance of the global model in non-IID data environments without compromising defense capabilities. Specifically, it utilizes UMAP to capture non-linear features among high-dimensional local models. Based on these features, robust regression and unsupervised clustering techniques are applied to effectively detect and remove attackers from local model updates. Subsequently, the server employs information stored in weights to evaluate and aggregate the remaining divergent model updates, thus significantly improving the global models performance. To validate the efficacy of Guard-FL, we provide a theoretical analysis of its convergence properties. Our experiments demonstrate that Guard-FL surpasses existing stateof-the-art solutions, achieving up to 96% accuracy in detecting malicious clients on non-IID CIFAR-10 datasets under various Byzantine attack scenarios. The implementation code is provided at https://github.com/XidianNSS/Guard-FL.git

Federated learning (FL) has gained popularity in the field of machine learning, which allows multiple participants to collaboratively learn a highly-accurate global model without exposing their sensitive data. However, FL is susceptible to poisoning attacks, in which malicious clients manipulate local model parameters to corrupt the global model. Existing FL frameworks based on detecting malicious clients suffer from unreasonable assumptions (e.g., clean validation datasets) or fail to balance robustness and efficiency. To address these deficiencies, we propose FedDMC, which implements robust federated learning by efficiently and precisely detecting malicious clients. Specifically, FedDMC first applies principal component analysis to reduce the dimensionality of the model parameters, which retains the primary parameter feature and reduces the computational overhead for subsequent clustering. Then, a binary tree-based clustering method with noise is designed to eliminate the effect of noisy points in the clustering process, facilitating accurate and efficient malicious client detection. Finally, we design a self-ensemble detection correction module that utilizes historical results via exponential moving averages to improve the robustness of malicious client detection. Extensive experiments conducted on three benchmark datasets demonstrate that FedDMC outperforms state-of-the-art methods in terms of detection precision, global model accuracy, and computational complexity.

Federated learning (FL) is vulnerable to poisoning attacks, where malicious clients tamper their model parameters to deteriorate the global model. Existing methods for defending against poisoning attacks primarily rely on identifying malicious clients, but struggle to balance robustness and efficiency. To address these issues, we propose FedPTA, a Prior-based Tensor Approximation (PTA) method. The core idea of FedPTA is to detect malicious clients in federated learning by leveraging inherent priors. This method initially innovatively defines multi-round model parameters as a three-dimensional tensor and unfolds it along different dimensions. Subsequently, three inherent priors - the similarity among benign clients, the continuity of multi-round client model parameters and the sparsity of malicious parameters, are integrated into a convex optimization framework. Through the optimization process, the optimal solutions for the background tensor and anomaly tensor are solved. Ultimately, the anomaly tensor is used to highlight the element-level features of malicious parameters, effectively distinguishing malicious clients. Evaluative studies supported by theoretical significance demonstrate the effectiveness of FedPTA, outperforming current state-of-the-art methods in terms of detection accuracy and computational efficiency.

Multidimensional mean estimation with local differential privacy (LDP) extracts the numerical features from groups while protecting users’ personal information without relying on a trusted server. However, the increase of dimensionality would lead to a deficiency in the allocable privacy budget, resulting in excessive accuracy loss. To solve this problem, we propose HyperMean, an effective privacy-preserving mean estimation mechanism for multidimensional data whose accuracy is at least no worse (and better in most cases) than existing solutions. We first design a multidimensional staircase function to obfuscate users’ data, significantly reducing the output variance. Second, an adaptive dimensionality reduction is performed on users’ data to allocate the privacy budget to some focused dimensions. Finally, the server averages all users’ obfuscated outputs to obtain an unbiased estimate of the mean results. Theoretical analysis reveals that HyperMean effectively reduces the worst-case variance of multidimensional mean estimation under LDP while maintaining low computational complexity. Experiments on both simulated and real-world datasets show that HyperMean outperforms existing multidimensional mean estimation mechanisms in terms of aggregated error.

Federated learning (FL), a cooperative distributed learning framework, has been employed in various intelligent Internet of Things (IoT) applications (e.g., smart health-care, smart home, and smart industry). However, there may be malicious devices in these IoT applications inferring other devices’ privacy or destroying the uploaded model parameters. Besides, due to the heterogeneity of IoT devices, it is difficult for the existing synchronized FL to effectively train models through non-identical independently distributed (non-IID) local data sets. To address these issues, we propose an asynchronous grouped federated learning framework (PAG-FL) for IoT, enabling multiple devices and the server to collaboratively and efficiently train models without revealing privacy. PAG-FL framework consists of an adaptive Rényi Differential Privacy-based privacy budget allocation (ARB) protocol and an asynchronous weight-based grouped update (AWGU) algorithm. In particular, our ARB protocol applies Rényi Differential Privacy and adaptively adjusts the privacy budget to obtain an efficient local model. The AWGU algorithm can defend against the poisoning attack on non-IID data set by weighing grouped local models to generate a global model. Meanwhile, it also realizes the asynchronous optimized update by adopting a lazy loading strategy. Theoretically, the proposed framework has a convergence guarantee and a privacy guarantee when training over the non-IID data set in an asynchronous FL. Our empirical experiments validate the effectiveness of the theoretical design and demonstrate the improved utility and robustness of PAG-FL in heterogeneous IoT.

Mobile healthcare (mHealth) enables people to collect and share their personal health records (PHRs) and gain rapid medical treatment via mobile 5G-enabled Industrial Internet of Things (IIoT) devices, which also brings the challenge of keeping the PHRs confidentiality and preventing unauthorized access. By the emerging ciphertext-policy attribute-based encryption (CP-ABE), the PHR owner can encrypt his/her PHR data under self-defined access policies. However, existing CP-ABE schemes are suffering from either heavy computation cost and storage overhead or traitor tracing and direct revocation. In this article, we propose an efficient, traceable, and revocable access control scheme named TRAC for mHealth in 5G-enabled IIoT. In TRAC, the ciphertext is composed of the attribute-relevant ciphertext encrypted under anand-gate access structure and the identity-relevant ciphertext associated with some potential receivers. The malicious user who leaks his/her privilege to unauthorized entities will be precisely tracked and added in the revocation list, by which the cloud server can update the identity-relevant ciphertext by itself. The length of final ciphertext and the time of bilinear pairing operations used in decryption are constant. The security analysis and performance evaluation indicate the security, efficiency, and practicality of TRAC.

Mobile crowdsensing systems use the extraction of valuable information from the data aggregation results of large-scale IoT devices to provide users with personalized services. Mobile crowdsensing combined with edge computing can improve service response speed, security, and reliability. However, previous research on data aggregation paid little attention to data verifiability and time sensitivity. In addition, existing edge-assisted data aggregation schemes do not support access control of large-scale devices. In this study, we propose a time-sensitive and verifiable data aggregation scheme (TSVA-CP-ABE) supporting access control for edge-assisted mobile crowdsensing. Specifically, in our scheme, we use attribute-based encryption for access control, where edge nodes can help IoT devices to calculate keys. Moreover, IoT devices can verify outsourced computing, and edge nodes can verify and filter aggregated data. Finally, the security of the proposed scheme is theoretically proved. The experimental results illustrate that our scheme outperforms traditional ones in both effectiveness and scalability under time-sensitive constraints.

Multiauthority ciphertext-policy attribute-based encryption (MA-CP-ABE) is a promising technique for secure data sharing in cloud storage. As multiple users with same attributes have same decryption privilege in MA-CP-ABE, the identity of the decryption key owner cannot be accurately traced by the exposed decryption key. This will lead to the key abuse problem, for example, the malicious users may sell their decryption keys to others. In this paper, we first present a traceable MA-CP-ABE scheme supporting fast access and malicious users’ accountability. Then, we prove that the proposed scheme is adaptively secure under the symmetric external Diffie–Hellman assumption and fully traceable under the q -Strong Diffie–Hellman assumption. Finally, we design a traceable and revocable MA-CP-ABE system for secure and efficient cloud storage from the proposed scheme. When a malicious user leaks his decryption key, our proposed system can not only confirm his identity but also revoke his decryption privilege. Extensive efficiency analysis results indicate that our system requires only constant number of pairing operations for ciphertext data access.

In cloud‐assisted electronic health care (eHealth) systems, a patient can enforce access control on his/her personal health information (PHI) in a cryptographic way by employing ciphertext‐policy attribute‐based encryption (CP‐ABE) mechanism. There are two features worthy of consideration in real eHealth applications. On the one hand, although the outsourced decryption technique can significantly reduce the decryption cost of a physician, the correctness of the returned result should be guaranteed. On the other hand, the malicious physician who leaks the private key intentionally should be caught. Existing systems mostly aim to provide only one of the above properties. In this work, we present a verifiable and traceable CP‐ABE scheme (VTCP‐ABE) in eHealth cloud, which simultaneously supports the properties of verifiable outsourced decryption and white‐box traceability without compromising the physician’s identity privacy. An authorized physician can obtain an ElGamal‐type partial decrypted ciphertext (PDC) element of original ciphertext from the eHealth cloud decryption server (CDS) and then verify the correctness of returned PDC. Moreover, the illegal behaviour of malicious physician can be precisely (white‐box) traced. We further exploit a delegation method to help the resource‐limited physician authorize someone else to interact with the CDS. The formal security proof and extensive simulations illustrate that our VTCP‐ABE scheme is secure, efficient, and practical.

In service-oriented mobile social networks (S-MSN), many location-based services are developed to provide various applications to social participants. Services can in turn be composed with the help of these participants. However, the composite structure, the subjective interpretation of trust demand, and the opportunistic connectivity make service composition a challenging task in S-MSN. In this paper, we propose a novel approach to enable trustworthy service evaluation and invocation during the process of composition. By analyzing dependency relationships, our approach can decentralizedly evaluate the trust degree of each service based on a lattice-based trust model to prevent data from being transmitted to untrustworthy counterparts. Besides, service consumers and vendors are able to specify their global and local constraints on the trust degree of service components on demand for more effective composition. Finally, by introducing acquaintances to the neighbors iteratively, social participants form a trust-aware acquaintance graph to forward invocation messages.

Attribute‐based encryption (ABE) has been considered as a promising cryptographic primitive for realising information security and flexible access control. However, the characteristic of attributes is treated as the identical level in most proposed schemes. Lattice‐based cryptography has been attracted much attention because of that it can resist to quantum cryptanalysis. In this study, lattice‐based threshold hierarchical ABE (lattice‐based t ‐HABE) scheme without random oracles is constructed and proved to be secure against selective attribute set and chosen plaintext attacks under the standard hardness assumption of the learning with errors problem. The notion of the HABE scheme can be considered as the generalisation of traditional ABE scheme where all attributes have the same level.

Dynamic service composition in wireless environment provides us with a promising approach to build complex applications based on the basic value-added services. In different network domains, multiple services may provide data with different security levels. In order to prevent from information leakage, information flow security is a major concern in composite services. However, the energy-limited nature of user terminal in mobile computing environments poses a significant challenge for the centralized information flow verification where the verification node need cost lots of computation and network resources. In this paper, we specify the security constraints for each service participant to secure the information flow in service chain based on the lattice model, and then present a decentralized information flow verification framework that cooperates different service participants to complete the verification process distributively with respect to their information flow policies. Through the experiments and evaluations, the results show it decreases the verification cost on single service node.

Time constrained service composition raises several problems. Researches on QoS-driven service composition provide some preliminary solutions, but there are still some unsolved issues, which can be attributed to the following reasons: (1) the huge time consumption of inter-domain validation, (2) the dynamic execution time of services and (3) the difficulty in defining time constraint due to the opaque feature of composite services. In this paper, we propose a novel service composition algorithm, which models the service composition as multi-domain scheduling problem with minimal service resources and time constraint. Each service is modeled as an exclusive resource during its execution period. By computing the inter-domain communications and available services in each domain, the domain with optimal utilization rate is obtained to arrange services. Meanwhile, loop parallelization is adopted when a service cannot be executed on schedule. Moreover, redundant services of the initial composition are further optimized. Our experiment results show that our approach can effectively achieve service composition with time constraint.