By synergizing external knowledge with generative capabilities, RAG is emerging as an important approach for enhancing the professionalism and reliability of Large Language Models (LLMs). However, due to the difficulty of fully ensuring the reliability of external knowledge sources, RAG systems are vulnerable to data poisoning attacks. Existing poisoning attacks are typically designed for a single specific query, exhibiting limited generalization capability. In addition, existing methods craft poisoned texts with low retrieval probability, leading to limited attack effectiveness. What is more, they also suffer from inefficiency in poisoned text construction. In this paper, we propose AutoTPA, a novel RAG poisoning attack, which targets all user queries containing the specific trigger. AutoTPA integrates dynamic query clustering, efficient and minimal candidate token set construction, as well as heuristic token evaluation algorithm, enabling automated and efficient construction of effective poisoned texts. All the attacker has to do is to determine the trigger to attack. Extensive experimental results demonstrate that AutoTPA achieves an attack success rate up to 95.5 % in less time while using only a token set sized at about 3.5 % of the original vocabulary. It significantly outperforms existing methods, showing clear advantages in both efficiency and effectiveness.

Static analysis presents significant challenges in alarm handling, where probabilistic models and alarm prioritization are essential methods for addressing these issues. These models prioritize alarms based on user feedback, thereby alleviating the burden on users to manually inspect alarms. However, they often encounter limitations related to efficiency and issues such as false generalization. While learning-based approaches have demonstrated promise, they typically incur high training costs and are constrained by the predefined structures of existing models. Moreover, the integration of large language models (LLMs) in static analysis has yet to reach its full potential, often resulting in lower accuracy rates in vulnerability identification. To tackle these challenges, we introduce BinLLM, a novel framework that harnesses the generalization capabilities of LLMs to enhance alarm probability models through rule learning. Our approach integrates LLM-derived abstract rules into the probabilistic model, using alarm paths and critical statements from static analysis. This integration enhances the model’s reasoning capabilities, improving its effectiveness in prioritizing genuine bugs while mitigating false generalizations. We evaluated BinLLM on a suite of C programs and observed 40.1% and 9.4% reduction in the number of checks required for alarm verification compared to two state-of-the-art baselines, Bingo and BayeSmith, respectively, underscoring the potential of combining LLMs with static analysis to improve alarm management.

Backdoor attacks against deep learning based object detectors have been studied increasingly in recent years. While most proposed attacks primarily focus on compromising the model’s integrity by inducing incorrect detections, only few studies explore backdoor attacks targeting the model’s availability, a critical concern in safety-critical domains such as autonomous driving. In this paper, we introduce a novel backdoor attack called the Sponge Backdoor Attack (SBA), designed to increase the detection latency of end-to-end object detectors. Specifically, we overload a commonly employed technique in many object detectors - non-maximum suppression (NMS) by introducing a large amount of non-existent objects. Through comprehensive experiments, we demonstrate the SBA’s effectiveness to prolong the processing time of the poisoned image while maintaining detection performance on clean images across various models, datasets, and hardware platforms.

The management of cryptographic keys within cloud services is crucial for ensuring data security but raises significant privacy and trust issues. This paper addresses the vulnerabilities of conventional key management systems (KMS), where cloud service providers might access and control user keys without consent, as well as intercept and misappropriate users’ secret data. To address these concerns, we propose a KMS protocol that leverages trusted hardware to minimize the level of trust required in cloud service providers. Our protocol ensures that the customer master key remains confined within a secure enclave, inaccessible to the service provider, and incorporates a bidirectional authentication mechanism to protect against impersonation by malicious providers. Within the universally composable framework, we design and analyze our protocol and rigorously prove its security. As a result, our designed KMS can be used in composition with any system, and its security can be guaranteed under any conditions. We also develop a prototype KMS based on Intel Trust Domain Extensions and evaluate its performance, emphasizing the system’s viability in real-world applications. The source code for our prototype is publicly available.

Android has designed Parcel as its high-performance serialization mechanism to pass objects across processes. For classes to be serialized by Parcel, developers must implement the methods for writing and reading the object's properties to and from a Parcel container. The inconsistency between those methods implemented by careless developers introduces Parcel Mismatch bugs, often occurring in vendor-customed classes due to lack of public scrutiny. Parcel Mismatch bugs can be abused by malicious applications to gain system privilege. However, no mature solutions exist to detect Parcel Mismatch bugs. This paper proposes PMDET, a fuzzing-based detection tool for Parcel Mismatch bugs. PMDET is capable of handling different vendors' firmware without actual devices. It loads Parcelable classes from Android firmware, emulates the Android runtime environment for Parcel to work, and monitors the serialization and deserialization procedures for mismatches. We evaluate PMDET with various Android firmware from different vendors. PMDET has identified 12 previously undisclosed mismatches, 6 of which are exploitable. Source code: https://github.com/tkmikan/pmdet.

Open social network (OSN) plays a more significant role in information propagation through the rapid developing of information technology. Since information diffusion is an essential process happens in OSN, it has been studied in many studies. Several models have been proposed to infer the diffusion process and reproduce diffusion network. However, these methods have two critical problems: 1) ignoring the effects of user social characteristics and 2) inaccuracy resulted from calculating the influence of different features independently. To address these limitations, a diffusion inferring method based on a recommender system (DIM-SPTF) was proposed. The DIM-SPTF method considers the propagation process between the users as the recommendation process of information and employs a recommender system to infer the propagation relationship. Through determining the propagation relations among all users in the observed topic data set, an information diffusion network can be finally obtained. Experimental results show that DIM-SPTF leads to improvements in performance compared with the state-of-the-art methods.

Privacy issues are concerned when data holders share their detected security data for correlation and analysis purpose. This paper proposes an approach to correlate and analyze intrusion alerts, while preserve privacy for alert holders. The raw intrusion alerts are protected by improved k-anonymity model, which preserves the alert regulation inside disturbed data records. With this privacy preserving technique, combing the typical FP-tree association rules mining algorithm, the approach provides the capacity of well balancing the alert correlation and the privacy preservation. Experimental results show that this approach works comparatively efficient and reaches a well balance between the alerts correlation and the privacy issues.