Retrieval Augmented Generation (RAG) enhances large language models (LLMs) by integrating external knowledge, but its open architecture exposes it to covert poisoning attacks. However, existing attacks often treat retrievability and textual naturalness as separate objectives, making them vulnerable to basic surface-level defenses. To address this limitation, we formalize a practical gray-box threat model and propose CPA-RAG, a framework that coordinates three attack conditions, retrievability, deceptiveness, and concealment, through iterative generation, rewriting, evaluation, and pruning. Using shadow retriever feedback to guide multi-model generation and iterative refinement, CPA-RAG maintains high concealment while effectively manipulating the retrieval and generation pipeline. Our evaluation demonstrates that CPA-RAG achieves an average attack success rate of 0.89 across three benchmark datasets and eight diverse LLMs. Under combined defense scenarios, it outperforms state-of-the-art black-box baselines by 14.5 percentage points. Furthermore, simulation experiments on commercial platforms confirm that CPA-RAG can affect generated outputs in these platform-level settings, with success rates between 91% and 97%. In accordance with ethical research practices, we have disclosed our findings to all affected LLM vendors. These results highlight the urgent need for more robust RAG-specific defenses.
Graph Neural Networks (GNNs) are increasingly deployed, making ownership verification critical for protecting their intellectual property against model theft. Fingerprinting and black-box watermarking are two main meth ods. However, the former relies on determining model similarity, which is computationally expensive and prone to ownership collisions after model post-processing. The latter embeds backdoors, exposing watermarked mod els to the risk of backdoor attacks. Moreover, both previous methods enable ownership verification but do not convey additional information, limiting verification efficiency. To address these challenges, this paper proposes WGLE, a novel black-box watermarking paradigm for GNNs that enables embedding the multi-bit string in GNN models without using backdoors. WGLE builds on a key insight we term Layer-wise Distance Difference on an Edge (LDDE), which quantifies the difference between the feature distance and the prediction distance of two connected nodes in a graph. By assigning unique LDDE values to the edges and employing the LDDE sequence as the watermark, WGLE supports multi-bit capacity without relying on backdoor mechanisms. We evaluated WGLE on six public datasets across six mainstream GNN architectures. Compared to state-of-the-art GNN watermark ing and fingerprinting methods, WGLE achieves perfect ownership verification accuracy with marginal fidelity degradation and exhibits robust resilience against potential ownership attacks.
Diamond demonstrates significant application potential in electronic devices owing to its excellent properties, yet its practical application is often limited by interface challenges in precisely controlling its interfacial electronic properties. This study demonstrates that surface termination engineering serves as a potential strategy for effectively modulating the electron affinity (EA) of diamond, thereby tuning the band offsets and Schottky barrier heights (SBHs) at its interfaces. By constructing 24 distinct diamond termination models, the EA of diamond can be tuned from -3.95 to 2.71 eV, governed by the combined effects of surface dipole moment and gap states. Furthermore, the terminated diamond forms a charge-blocking layer at the heterojunction interface, thereby largely preserving the original band positions. Consequently, a linear modulation of the band offset from 0.58 to 4.34 eV is accomplished by integrating beta-Ga2 O3 with 12 different diamond terminations. For metal/diamond contacts, an SBH adjustment from -0.78 to 2.14 eV is achieved at Pt/diamond interfaces by employing different terminations. Notably, the introduction of surface-state-free terminations effectively suppresses metal-induced gap states, thereby achieving significant Fermi-level depinning. This work elucidates the underlying mechanism of termination-induced EA tuning and establishes a paradigm for quantitative band offset and SBH control, providing critical guidance for the rational design of high-performance diamond-based devices. (c) 2026 Published by Elsevier Ltd on behalf of The editorial office of Journal of Materials Science & Technology.
Text-to-image diffusion models have recently demonstrated significant progress in video editing. However, existing video editing methods are severely limited by their high computational overhead and memory consumption. Furthermore, these approaches often sacrifice visual fidelity, leading to undesirable temporal inconsistencies and artifacts such as blurring and pronounced mosaic-like patterns. To address these dual challenges and strike a balance between computational efficiency and visual fidelity, we propose Edit-Your-Interest, a lightweight, text-driven, zero-shot video editing method. Edit-Your-Interest introduces a Spatio-Temporal Feature Memory bank (SFM) to cache spatial-attention features from previous frames under a fixed memory budget. The SFM maintains a representative candidate pool, reducing the computational overhead of full-sequence spatio-temporal modeling. To query this memory, we propose Feature Most-Similar Propagation (FMP), a similarity-based retrieval rule that propagates relevant cached tokens to the current frame. Together, SFM and FMP form a bounded-memory retrieval-based propagation design that improves temporal consistency while reducing blurring and mosaic-like artifacts. Finally, we introduce a feature-diversity-based SFM update algorithm to keep the memory representative and non-redundant throughout the video sequence. Furthermore, to enable precise object editing, we leverage cross-attention maps to automatically extract masks for the instances of interest. These masks are seamlessly integrated into the diffusion denoising process, enabling fine-grained control over target objects and allowing Edit-Your-Interest to perform highly accurate edits while robustly preserving the background integrity. Extensive experiments demonstrate that Edit-Your-Interest outperforms state-of-the-art methods in both efficiency and visual fidelity. The project page is available at https://edit-your-interest.github.io.
This paper presents an automated framework for sensor insertion attacks detection in scalable dis crete event systems modeled by enhanced Petri nets. Insertion attack represent a significant threat to modern systems, particularly in large-scale and complex infrastructures. In such scenarios, at tackers may falsely report events that have not occurred, causing observers to make incorrect decisions. The proposed framework combines Petri net modeling with observation-augmented structures to generate and analyze an extended reachability graph, while using integer linear pro gramming to obtain identification places for accurate attack detection. This process removes the need for manual intervention and reduces the impact of the state explosion problem. The generated markings are classified as positive or negative, where negative markings indicate the occurrence of an attack. The framework relies on well-defined mathematical formulations, including formal definitions, theorems, and algorithms, to provide analytical rigor and computational efficiency. Experimental results show that the framework is scalable and performs effectively across different system configurations. These results demonstrate its suitability as a practical solution for securing large-scale cyber-physical systems.
