Large Language Models (LLMs) have recently gained popularity for tasks like text generation and automated reasoning, often hosted on cloud servers to handle large-scale user interactions. Yet this setup faces a critical challenge, prompt injection attacks, where adversaries craft malicious text to bypass an LLM’s safety constraints and policy filters. Attackers constantly evolve new injection tactics, rendering static defenses ineffective. While service providers may wish to store user inputs for security audits, this practice poses serious privacy and data-protection risks. User prompts can contain personal or confidential information subject to strict regulatory requirements. To address these conflicting demands of security monitoring and user privacy, we propose a privacy-preserving prompt injection detection scheme via hash and embedding-based auditing for smart cloud-deployed LLMs. Our approach segments each user prompt into smaller chunks and stores only non-invertible representations of them—cryptographic hashes for exact matches and vector embeddings for near-match detection. This enables robust identification of both verbatim and paraphrased threats without retaining any plaintext data. Furthermore, newly discovered injection patterns can trigger retrospective audits on previously stored chunks, greatly enhancing attack traceability. We also introduce an optional blockchain-based logging mechanism for tamper-proof event records. Empirical results show that our method achieves high detection accuracy against diverse injection techniques while safeguarding sensitive user information.
With the growing integration of Smart Microgrids, cybersecurity concerns have become increasingly critical. Micro-grids face a variety of sophisticated cyber threats, including False Data Injection Attacks (FDIA), Denial of Service (DoS) Attacks, and Privilege Escalation Attacks, which compromise secure operations by obstructing communication networks and control systems. Existing security architectures lack the flexibility and adaptability to address the unique demands of microgrids. This paper proposes a novel Key Distribution Security Association-Based Confidentiality Isolation Control Framework, leveraging a dynamic GDOI multicast security association allocation method. The framework introduces fine-grained cryptographic boundary definitions and secure interconnection mechanisms to enhance microgrid security. By employing distinct Security Parameter Indexes (SPI) and integrating heartbeat signals, the framework achieves independent Security Associations (SA), effectively addressing issues like source authentication, replay attacks, and resource inefficiencies. Simulations and security analyses validate the proposed framework’s ability to defend against FDIA and replay attacks while reducing storage overhead and supporting dynamic adaptability in complex environments. This framework provides a scalable and practical solution for the secure operation of Smart Microgrids.
The rapid advancement of sixth-generation (6G) wireless networks has accelerated the adoption of terahertz (THz) communication due to its ultra-high data rates and abundant spectrum. However, distance estimation in non-line-of-sight (NLoS) THz environments remains challenging due to severe path loss, molecular absorption, and multipath effects. Moreover, precise localization raises privacy concerns, exposing nodes to adversarial tracking. To address these challenges, we propose a novel blind-distance estimation with adaptive privacy-preservation scheme for non-line-of-sight THz-based GPS-denied networks. We propose a novel blind-distance estimation method using a single THz frequency without requiring prior system knowledge. Second, we formulate a distance estimation model that integrates molecular absorption effects and multipath propagation to improve distance estimation accuracy. Third, we design a novel adaptive privacy-preserving perturbation mechanism that dynamically adjusts noise based on multipath propagation and obstacle density, ensuring feasible perturbations while balancing privacy and localization accuracy. Simulation results demonstrate that our scheme effectively balances localization accuracy and privacy preservation, adapting to both ideal and NLoS conditions in structured (urban) and unstructured (disaster) environments.
The integration of federated learning and recommendation systems is emerging as a prominent trend in machine learning, enabling personalized recommendations while preserving user privacy. In this paradigm, a master model is distributed to users, and the users perform local updates using their private data. The updates are sent back and aggregated on the server to update the master model then redistributed to the users. However, traditional federated recommendation systems encounter serveral challenges, including potential privacy leakage and high communication costs. To address these issues, we propose SIGN-FCF, a federated recommendation method based on matrix factorization, which leverages the SIGNSGD algorithm and differential privacy techniques. The proposed method employs a 1-bit compressor to enhance privacy protection and reduce communication costs, and three instances of the compressor are created to meet various privacy requirements. We also evaluate the performance of SIGN-FCF on three real-world datasets, demonstrating its effectiveness in preserving user privacy without compromising accuracy.
Many organisations utilise cloud-based data warehouses for their storage and analytics because these systems support multiple user tenants yet present considerable privacy threats. Traditional centralised data mining devices are sensitive to security risks, thus making federated learning (FL) an attractive solution for protecting information during mining processes. A research investigation analyses FedAvg and FedProx through a federated system while adding DP and secure aggregation technology to address security requirements. Tests demonstrate that SF-XGBoost delivers 99.31% accuracy alongside 0.9994 ROC-AUC values, which exceeds the other studied models in both centralised and federated learning platforms. The stability performance of FedProx is higher than that of FedAvg, although its precision and recall capabilities fall behind. Patronage clients that applied differentially private SF-XGBoost recorded 73.75% data accuracy while displaying privacy-utility configuration aspects. The slightly more significant losses of federated models parallel their ability to protect data privacy while satisfying privacy regulations. SF-XGBoost and privacy-improving methods deliver an effective security solution for cloud-based federated learning applications. Scientists must investigate how homomorphic encryption works alongside adaptive privacy budgets and federated transfer learning to achieve maximum optimisation of federated models.
