计算机科学最新文献

Drug-drug interaction (DDI) prediction seeks to identify pharmacodynamic and pharmacokinetic interactions arising from the co-administration of multiple drugs with differing physicochemical properties. However, most existing methods predominantly focus on global molecular or atomic topologies, overlooking the critical role of dynamically derived functional substructures (motifs) in modulating drug activity and the occurrence and strength of DDI. To this end, we propose a dissimilarity-driven dual-interaction via co-occurrence motifs and complementary fragments (D3I-COMCF) framework that deeply integrates molecular-, motif-, and atomic-level information for accurate DDI prediction. Specifically, we construct, for the first time, a co-occurrence motifs molecular interaction graph (CMMIG) that unifies the global molecular structure, local physicochemical properties of motifs, and molecule-motif interaction topology. By weighting rare, dissimilar but information-rich motifs, D3I-COMCF was guided to focus on more discriminative functional units. Moreover, based on the maximum common substructure mapping and its dissimilar one-hop neighbor extension, we dynamically incorporated complementary fragments attached to the shared backbone into a fragment interaction graph, revealing electronic complementarity and spatial cooperativity with computational efficiency. Finally, hierarchical multi-block integration of the three levels enables cross-scale joint representation of drug pairs. Extensive experiments on multiple public datasets verified the effectiveness of D3I-COMCF, which achieved strong and stable performance under standard benchmark protocols and showed promising cold-start generalization and interpretability. Additional hard-negative analysis further indicates that D3I-COMCF remains sensitive to known positive DDIs, while controlling false positives among structurally similar candidate pairs not annotated as known positives remains challenging.

Unsupervised video anomaly detection aims to learn normal spatio-temporal patterns from unlabeled surveillance data and identify deviations during inference. Existing methods often struggle to capture long-term temporal dependencies and adapt to spatial variations caused by target deformation and viewpoint changes. Although diffusion-based models provide stable distribution modeling, their limited temporal conditioning and fixed spatial representations restrict anomaly discrimination in complex scenes. To address these issues, we propose a boundary regularization enhanced video anomaly detection network based on context-adaptive spatio-temporal conditional diffusion. Rather than relying on rigid feature extraction, our method explicitly models long-range motion evolution and dynamically adapts to spatial variations caused by target deformation and viewpoint shifts. Furthermore, to prevent overfitting to normal behaviors, we introduce a hybrid pseudo-anomaly generation strategy that acts as a boundary regularization constraint, explicitly sharpening the model's discriminative capability against abnormal patterns. Experiments on benchmark datasets including UCSD Ped2, CUHK Avenue, and ShanghaiTech demonstrate that our method significantly improves detection accuracy. By effectively integrating context-adaptive conditioning and boundary regularization, this work establishes a highly discriminative and generalizable paradigm for video anomaly detection, demonstrating strong robustness in complex, real-world surveillance scenarios.

Maintaining a precise remote center of motion is essential for safe and accurate tool manipulation in retinal microsurgery. However, existing numerical Jacobian identification methods for parallel manipulators often exhibit nonlinear, workspace-dependent inaccuracies and require frequent recalibration, limiting their reliability and clinical applicability. To address these challenges, this study presents an analytical kinematic framework for a hybrid parallel-serial robot designed for retinal surgery, known as the Steady-Hand Eye Robot (SHER 3.0). Closed-form solutions for forward and inverse kinematics, as well as analytical formulations of the direct and inverse Jacobians, are derived to ensure consistent motion estimation across the workspace. The kinematic performance of SHER 3.0 is analyzed to evaluate the manipulability and workspace efficiency. Building on these models, a model predictive control strategy is implemented on SHER 3.0, which maintains the remote center of motion constraint at the sclerotomy with sub-millimeter accuracy, with a root mean square error of 0.55 ± 0.12 mm, in a pilot study on a teleoperation experiment. Experimental results demonstrate the effectiveness of the proposed analytical models and control framework in enhancing robot motion stability and robustness of the controller in maintaining the remote center of motion constraints, enabling safe robot-assisted retinal microsurgery.

Various defenses have been proposed to counter backdoors in deep neural network classifiers, but due to the diverse mechanisms through which backdoor triggers can be incorporated into data samples, no single defense is universally effective. We introduce an ensemble defense that consolidates the complementary strengths of individual detectors, making the detection system more backdoor mechanism-agnostic, and also providing backdoor inversion, which offers insights into the nature of the detected attack. Our ensemble defense achieves superior detection performance and more robust, accurate backdoor inversion compared to individual detectors. We also employ mixed clean/dirty-label backdoor poisoning, a recent X-to-X backdoor attack that is more surgical, evasive, and harder to detect than traditional dirty-label attacks.

The increasing prevalence of advanced persistent threats and automated cyberattacks highlights the need for proactive cybersecurity strategies. Reinforcement-learning-based autonomous penetration testing has shown promise; however, many existing simulation environments rely on abstract, probability-driven attack models, resulting in limited reproducibility and gaps relative to real-world conditions. This paper presents E-NASim as a reinforcement-learning-based cyberattack simulation framework that explicitly models attack feasibility and system evolution using pre-/post-attack state (PPS) transitions derived from MITRE ATT&CK subtechniques. In E-NASim, attack execution conditions and state transitions are deterministically governed by PPS definitions, whereas reinforcement learning is used to learn effective multistage attack sequencing under fixed constraints. Experimental results demonstrate that E-NASim enables agents to learn coherent and executable multistage attack strategies in complex network environments, providing a practical foundation for automated penetration testing and security evaluation.

Traditional knowledge distillation transfers information from a large teacher model to a compact student by aligning static output distributions or internal representations for fixed inputs. Such formulations overlook how predictions change in response to input perturbations, even though robustness is critical for deployment. We interpret distillation as the transfer of functional response behavior under input variations and propose cross-corruption distillation (CCD), which aligns corruption-induced differences in teacher and student output via a Kullback–Leibler objective. Rather than matching absolute logits, CCD distills how the teacher redistributes its prediction mass under perturbations, transferring teacher-conditioned input sensitivity without architectural correspondence. On diverse multimodal benchmarks, CCD performs competitively, often outperforming conventional logit-based distillation. It also slows performance degradation as perturbation strength increases and remains stable across corruption strategies and training schedules. These results position CCD as a response-aware distillation framework integrating robustness directly into the distillation objective.

Safety alignment in large language models is overwhelmingly trained on English data, yet models are deployed across typologically diverse languages. We evaluate six open-weight large language models (LLMs) across 17 bilingual Korean/Chinese prompt conditions and three translation engines (NLLB-200, Gemini 3.0 Pro, and GPT-5.2), generating 53 040 AdvBench responses. We also conduct French/Spanish and HarmBench robustness checks. The regex classifier is cross-validated against an open-weight LLM judge (Qwen-2.5-72B, $�\sim$ 7500 judgments), and we report Wilson 95% confidence intervals. Three findings hold for both classifiers: Safety erosion is language-specific, not generically non-English; Chinese-origin models (Yi, InternLM) exhibit asymmetric vulnerability; and a two-tier pattern persists, where models with target-language safety data resist high-quality translations, and models without such safety data remain vulnerable, regardless of translation quality. These findings reframe cross-lingual safety as a per-language training coverage problem.

Large language models are vulnerable to prompt injection because transformer attention treats tokens too uniformly across system and user roles. We propose the adaptive orthogonal role-aware transformer (AORT), a lightweight retrofit that improves role-sensitive decoding. AORT preserves the base token embeddings and adds at each transformer layer a learned orthonormal re-encoding and a role-gated additive attention bias to maintain separation between instruction and user representations during context mixing. We evaluate AORT on direct and indirect prompt-injection benchmarks, including the BIPIA-text benchmark. AORT achieves the lowest attack success rate (ASR) in most settings and successfully resists all eight evaluated attacks on LLaMA 3.1–8B. AORT reduces the ASR from 13.6% to 2.1% on LLaMA 3.1–8B evaluated with the BIPIA-text benchmark. These findings indicate that explicit role-priority structure in the forward pass can substantially improve prompt-injection robustness with minimal architectural overhead.