Retrieval-Augmented Generation (RAG) systems enhance large language models (LLMs) by retrieving relevant external documents, often from proprietary and sensitive private databases, to improve response accuracy. However, recent studies have revealed significant privacy risks to these databases in the form of Membership Inference Attacks (MIAs), which allow adversaries to determine whether specific documents are included in a RAG system's retrieval corpus through carefully crafted queries. Existing MIA approaches either rely on excessive queries, making them vulnerable to detection, or suffer from unsatisfactory inference accuracy. Achieving high attack performance under strict query constraints remains a challenging open problem. To address this challenge, we propose Few-MIA, a few shot MIA method tailored for RAG systems. Our key insight is that queries targeting member documents tend to consistently retrieve relevant document-specific chunks, whereas queries targeting out-of-corpus content (non-members) produce semantically inconsistent retrieved information, leading to observable differences in the generated responses. Few-MIA leverages this discrepancy by generating a small set of semantically consistent but syntactically diverse queries using an auxiliary LLM and inferring membership based on the consistency of the system's responses. We conduct extensive evaluations across three real-world datasets, four mainstream generators, and two widely-used retrievers. Experimental results show that Few-MIA significantly outperforms existing state-of-the-art (SOTA) approaches in both attack success rate (achieving an average of 94.5%) and query efficiency (a 7.5×improvement), and substantially improving stealth and robustness against advanced detection. Few-MIA achieves a good balance between attack effectiveness and query efficiency in a black-box setting.
Multimodal failure diagnosis is promising in maintaining microservice availability. However, it is challenging to acquire high-quality multimodal data in dynamic environments. When confronted with incomplete modalities, existing methods allow low-quality modalities to mislead the diagnostic process, leading to degraded performance. Prior reconstruction-based methods primarily target metrics, but are unsuitable for logs and traces due to prohibitive overhead and severe hallucinations caused by the irreversible data missingness. Furthermore, existing fusion strategies typically treat all modalities equally, blindly assuming equal reliability across them. In this paper, we propose MUSE, an uncertainty-aware multimodal fusion framework engineered to be robust against modality missingness. This framework transitions from blind fusion to multimodal quality sensing by leveraging predictive uncertainty, which is derived from energy-based logit distributions. MUSE introduces modality-specific encoders and a learnable dual-constraint optimization strategy, which ensures consistency between quality weights and cross-modal attention. Finally, MUSE achieves sample-level adaptive attention adjustment based on feature reliability. We evaluate MUSE on two widely used datasets from real world microservice systems. Experimental results demonstrate that MUSE consistently outperforms state-of-the-art baselines under complete modalities. Furthermore, across four missingness scenarios, MUSE effectively mitigates modality incompleteness, yielding average F1-score improvements of 7.85% and 5.43%.
The rapid growth of generative image technology has made deepfake images increasingly indistinguishable from real ones, leading to widespread abuse of generated images. This abuse inevitably stems from the underlying generative models, making it challenging to trace the origins of these deepfake images back to their source models. This paper proposes a proactive defense scheme for tracing deepfake models using undeniably efficient model fingerprinting from the developer's perspective. Apply artificial fingerprinting technology to preprocess the training images from the generative model's training dataset. By validating the transferability of fingerprints from training images to the model and then to generated images, it greatly simplifies deepfake image tracing into a fingerprint detection and matching task. By cryptographic commitment primitives during model training and fingerprint embedding, this approach ensures the undeniability of the original model fingerprint. The polynomial identity testing is employed to compute the polynomial values of the fingerprint, enabling lightweight and efficient verification through validation of the fingerprint polynomial. For instance, verifying a 100 bits fingerprint requires only 30 to 40 bits of communication, reducing overhead by 60 ∼ 70%. Security analysis and experiments demonstrate that the proposed fingerprint scheme reduces communication complexity to O(logL) (L denotes the fingerprint length), achieving unforgeable security and robust undeniability.
Distributed consensus in multi-agent systems (MASs) is critical for coordinated decision-making and widely applied in distributed computing, UAV, and smart grids. How ever, many methods in the past faced privacy leakage risks during communication processes, while existing approaches such as homomorphic encryption and differential privacy encounter challenges like high computational overhead or the inability to achieve precise consensus. To address this, this paper proposes a privacy-protected sub-state fusion algorithm (PSFA) based on noise addition and state decomposition. The algorithm adopts a dual privacy protection strategy: first, in secure initialization of noise variables, each agent generates a group of independent random signals via an encrypted channel, and constructs its local noise for the initial state of the agent; then, through the state decomposition mechanism, it is divided into multiple sub state values. By only exchanging the sub-states rather than the complete information, the risk of privacy leakage is effectively reduced. Furthermore, this paper introduces a hierarchical PSFA algorithm (HPSFA), which dynamically adjusts the number of sub-states based on agents' privacy requirements to balance privacy preservation and system efficiency. Theoretical analysis demonstrates that the introduction of zero-mean noise eliminates the statistical bias inherent in traditional differential privacy, enabling exact average consensus. Simulation results verify that the algorithm proposed in this paper can achieve the tradeoff be tween convergence speed and privacy protection, and successfully applied to the smart grid economic adjustment.
Network anomaly detection can provide good protection for network devices. As a mainstream detection algorithm, machine learning/deep learning based intrusion detection systems face the serious impact of imbalanced network traffic training data. A large number of oversampling algorithms have been proposed for the augmentation of network anomaly detection algorithms, but the following problems still exist. In binary classification, it is difficult to learn the complete distribution of attack categories based on the existing imbalanced training set. In multiclassification, there is a problem of class overlap on top of the difficulty of learning the complete distribution. In the case where the generated traffic can be replayed, it is difficult to find a solution in the problem space that can cope with the imbalance based on effective feature space augmentation samples. For these reasons, we propose the HeteAUG model. On feature space augmentation, we combine interpolation-based and convolutional generative adversarial networks-based augmentation results in a heterogeneous form. For problem space augmentation, a heuristic search algorithm with constraints is used on the basis of effective feature space augmentation samples. Experiments show that our augmentation model outperforms other state-of-the-art (SOTA) models. Specifically, taking the complex Kitsune dataset as an example and using accuracy after augmentation as an evaluation metric, in binary augmentation, we outperform the SOTA algorithm by at least 6.1%. In multiclassification augmentation, we outperform the SOTA algorithm by at most 12.1%. In the augmentation of the binary classification problem space, we compare the accuracy of the classifiers before and after augmentation, and we outperform the imbalanced state by 33.5%. In addition, we discuss the on-the-ground deployment of the model and demonstrate its feasibility.
