The increasing sophistication of AI-generated misinformation poses a significant challenge, particularly in Vietnamese digital environments, where scam tactics are evolving rapidly. In response, this paper presents FraudTrace, an automated, multi-agent verification system designed to detect and explain misinformation and online fraud. Our key contributions include (1) the construction of a diverse and realistic dataset comprising 4,221 labeled samples, (2) the introduction of a URC (Understandable Response Clarity) metric for evaluating model explainability, and (3) the deployment of a modular, role-based multi-agent architecture integrating fine-tuned Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), Google Search APIs, and scam detection frameworks. Experimental results indicate that fine-tuned LLMs significantly outperform zero-shot baselines, with the best-performing model achieving over 91
The rapid development of the Internet of Things (IoT) has realized the mutual communication between intelligent devices such as sensors, smart phones and vehicles. However, it is a challenge to design an authentication and key agreement protocol suitable for resource-constrained Industrial Internet of Things (IIoT) environment. Recently, Zahednejad et al. proposed a robust three-factor authentication protocol based on Elliptic Curve Diffie-Hellman (ECDH), and claimed that the protocol can resist various attacks. However, we found that their scheme could not resist sensor node capture attack. In order to overcome this problem, we propose an improved three-factor authentication scheme based on physical unclonable function (PUF). Informal analysis shows that the scheme has the ability to resist multiple attacks. In addition, we employed the Real or Random (ROR) model to conduct a formal analysis of the scheme, and the results indicate that the scheme is secure. Finally, compared with other similar schemes, this scheme has advantages in computation and communication costs, and is more practical.
As organizations navigate the complexities of the digital era, they face a growing and particularly insidious threat: internal cyber-attacks. These insider threats, whether intentional or unintentional, present unique cybersecurity challenges. This paper introduces a novel approach to real-time anomaly detection for insider threat hunting, leveraging sliding windows to analyze event streams at the finest granularity, event by event. Our methodology employs a range of unsupervised learning algorithms, including conventional machine learning methods, graph neural networks (GNN), and natural language processing (NLP) techniques, to identify anomalies without prior knowledge of potential attacks. Using sliding windows, we ensure continuous and contextual analysis, enabling real-time processing and detection. This approach is evaluated using the CERT insider threat dataset, renowned for its richness and believability, which simulates the activities of a large organization over an extended period, offers between 0.95 and 1 in usual ML metrics, and a true real-time process. By unifying these techniques under a shared framework, our solution aims to provide a robust and practical approach for the real-time detection of anomalies involved in insider threat activities, thereby enabling organizations to protect their critical assets from potential damage by malicious or negligent insiders.
With rising global energy consumption, smart grids require integrating massive, scattered device data, making energy-efficient data aggregation techniques critical. Yet, they face key challenges: balancing data privacy, boosting operational efficiency, and ensuring accurate anomaly detection across more scenarios. The article presents a distributed data aggregation frame for smart grids that guarantees data integrity amid malicious nodes or adversarial attacks. The system, governed by rigorous privacy protections, enables the rapid identification of data that surpasses specified criteria while maintaining the accuracy and uniformity of normal data through integrity verification mechanisms. This is, to our knowledge, the inaugural solution that concurrently incorporates malicious threat models and meter data integrity verification in distributed data gathering environments while preserving anomaly detection functionalities. Assessments of security, computational burden, and privacy safeguarding efficacy indicate that the suggested method attains an optimal balance among these elements, fully fulfilling the stringent demands of contemporary smart grid applications for both performance and security.
The diffusion model has gained considerable attention as an advanced deep generation model, showcasing remarkable performance. However, concerns have emerged regarding privacy and equity due to potential model misuse, especially for some sensitive features such as gender. Our focus is on addressing the challenge of machine unlearning within an unconditional diffusion model that selectively omits specific features. Our objective is to modify images generated by a pre-trained model by selectively removing particular image components. To achieve this, we propose a novel learning framework for unconditional image-to-image diffusion models. This framework integrates a scoring model into a pre-trained diffusion model, allowing for refinement without the need for a complete retraining process. Experimental validation using datasets such as MNIST and CelebA demonstrates the effectiveness of our approach. It highlights the model’s capability to generate and remove target features while maintaining the fidelity of the original image.
