arXiv subjects

Yinhao Xiao

Publications and source records attributed to Yinhao Xiao.

17 recordsLinked to original sources

arXiv2026

Mask2Shield: Strengthening LLM Safety against Neuron-Pruning Attacks

Large language models (LLMs) are safety-aligned before deployment to reduce harmful content generation. Yet neuron-level pruning attacks show that refusal can depend on a small set of removable units: disabling them can remove safety behavior while leaving much of the model usable. To address this problem, we introduce Mask2Shield (M2S), a masked-forward alignment method that trains a model under this functional pruning. The masked student must recover a safe refusal through the remaining computation, while a frozen, unmasked teacher supplies complete benign answers to limit capability drift. Across ten model configurations, M2S reduces successful recomputed pruning attacks from 80--279 to 1--44 out of 313 prompts while generally preserving four capability benchmarks. We also evaluate M2S with TwinBreak, which uses a different neuron-selection rule and iterative pruning procedure. Together, these results show that M2S makes targeted pruning less effective by reducing reliance on a small, removable safety-neuron set.

cs.CR↗

arXiv2026

Bridging Diffusion Pruning and Step Distillation with Teacher-Aligned Repair

Diffusion models generate high-quality images, but their inference cost comes from two sources: large denoising networks and repeated denoising steps. Existing compression pipelines usually attack these costs separately. Pruning reduces the network, but most pruning methods still rely on a long post-pruning retraining stage to recover a many-step sampler. Step distillation reduces the number of denoising steps, but it usually assumes a student that can already follow the teacher well enough to receive useful distillation gradients. This paper asks whether post-pruning retraining can be replaced by step distillation. We find that the direct replacement fails: after pruning an EDM2-XS teacher, starting SiDA from the pruned checkpoint produces unusable samples. We introduce a short teacher-alignment repair stage as a bridge between pruning and step distillation. The bridge matches the pruned generator to the teacher on noisy real-image latents, then hands the repaired checkpoint to one-step distillation. On ImageNet-512, the original EDM2-XS baseline uses 124.713M parameters and 63 network evaluations, reaching an FID of 3.53. With a suitable distillation objective, our 20% pruned one-step generator uses 98.826M parameters and one network evaluation, reaching an FID of 3.12. With 30% pruning, the model uses 88.029M parameters and one network evaluation, with an FID of 4.26.

cs.CV↗

arXiv2026

GeoContra: From Fluent GIS Code to Verifiable Spatial Analysis with Geography-Grounded Repair

Reliable spatial analysis in GIScience requires preserving coordinate semantics, topology, units, and geographic plausibility. Current LLM-based GIS systems generate fluent scripts but rarely enforce these geographic rules at scale. We present GeoContra, a verification and repair framework for LLM-driven Python GIS workflows. It represents each task as an executable geospatial contract-including natural-language questions, schemas, CRS metadata, expected outputs, spatial predicates, topology, metrics, required operations, and forbidden shortcuts. Generated programs undergo static rule inspection, runtime validation, and semantic verification, with violations fed back into a bounded repair loop. Evaluated on 7,079 real geospatial tasks across 15 Boston-area zones, 9 task families, and 11 open-source models (600 runs each), GeoContra improves spatial correctness on closed models from 47.6% to 77.5% for DeepSeek-V4 and from 57.7% to 81.5% for Kimi-K2.5. Across 11 open models, average correctness rises by 26.6%. GeoContra turns fluent code production into verifiable spatial analysis, catching negative travel times, CRS/field-schema violations, missing predicates, and brittle output casts that otherwise yield executable but geographically invalid results.

cs.SE↗

arXiv2026

Efficient Diffusion Distillation via Embedding Loss

Recent advances in distilling expensive diffusion models into efficient few-step generators show significant promise. However, these methods typically demand substantial computational resources and extended training periods, limiting accessibility for resource-constrained researchers, and existing supplementary loss functions have notable limitations. Regression loss requires pre-generating large datasets before training and limits the student model to the teacher's performance, while GAN-based losses suffer from training instability and require careful tuning. In this paper, we propose Embedding Loss (EL), a novel supplementary loss function that complements existing diffusion distillation methods to enhance generation quality and accelerate training with smaller batch sizes. Leveraging feature embeddings from a diverse set of randomly initialized networks, EL effectively aligns the feature distributions between the distilled few-step generator and the original data. By computing Maximum Mean Discrepancy (MMD) in the embedded feature space, EL ensures robust distribution matching, thereby preserving sample fidelity and diversity during distillation. Within distribution matching distillation frameworks, EL demonstrates strong empirical performance for one-step generators. On the CIFAR-10 dataset, our approach achieves state-of-the-art FID values of 1.475 for unconditional generation and 1.380 for conditional generation. Beyond CIFAR-10, we further validate EL across multiple benchmarks and distillation methods, including ImageNet, AFHQ-v2, and FFHQ datasets, using DMD, DI, and CM distillation frameworks, demonstrating consistent improvements over existing one-step distillation methods. Our method also reduces training iterations by up to 80%, offering a more practical and scalable solution for deploying diffusion-based generative models in resource-constrained environments.

cs.CV↗

arXiv2026

EvoPatch-IoT: Evolution-Aware Cross-Architecture Vulnerability Retrieval and Patch-State Profiling for BusyBox-Based IoT Firmware

BusyBox is one of the most widely reused userland components in Linux-based Internet-of-Things (IoT) firmware, yet its security assessment remains difficult because firmware images are frequently stripped, vendor patch practices are inconsistent, and the same source component is compiled for heterogeneous architectures. We propose EvoPatch-IoT, an evolution-aware cross-architecture retrieval framework for stripped BusyBox firmware binaries. EvoPatch-IoT combines anonymous instruction/context features, graph-level statistics, per-binary geometric priors, and historical function prototypes to localize homologous and potentially vulnerable functions without relying on symbols, source paths, or version strings at test time. We further construct a large-scale BusyBox benchmark from 57 historical versions, 270 unstripped binaries, 285 stripped binaries, and 130 source releases, yielding 1,550,752 function-symbol rows, 1,290,369 analysis-function rows, and 155,845 high-confidence stripped-to-unstripped matches. On 57 fully covered versions and 1,020 directed architecture pairs, EvoPatch-IoT achieves a weighted Hit@1 of 34.56\% and Hit@10 of 56.24\%, outperforming the strongest baseline by 16.04\% and 26.85\%, respectively, and reducing the expected manual inspection space by 98.98\%. The method is best on 56 of 57 versions and maintains consistent advantages on difficult architecture pairs. In addition, a version-change transfer study reaches a mean ROC-AUC of 0.9887, and a CVE-2021-42386 patch-state proxy obtains 82.44\% mean accuracy and 88.47\% mean F1 across held-out architectures. These results show that evolution-aware binary retrieval is a practical foundation for scalable IoT firmware vulnerability auditing.

cs.CR↗

arXiv2026

HWE-Bench: Can Language Models Perform Board-level Schematic Designs?

Large Language Models (LLMs) have demonstrated significant potential in various engineering tasks, including software development, digital logic generation, and companion document maintenance. However, their ability to perform board-level circuit design is understudied, as this task requires a synergized understanding of real-world physics and Integrated Circuit (IC) datasheets, the latter comprising detailed specifications for individual components. To address this challenge, we propose \hweb, an evaluation framework that benchmarks the ability of LLMs to perform such designs. It consists of 300 board-level design tasks pulled from open-source and crowdsourcing platforms such as GitHub and OSHWLab, covering 8 application domains, and is complemented with a knowledge base of 2,914 real IC datasheets. For each task, the LLMs are tasked with generating a schematic from scratch, using the provided circuit functional requirements and a set of component datasheets as input. The resulting schematic will be checked against a static electrical rules, and then passed to a circuit simulator to verify its dynamic behavior. Our evaluation show that although current models achieve initial engineering usability and documentation understanding, they lack physical intuition, as the top-performing model achieved an overall pass rate of 8.15\%. We envision that advancements on \hweb\ will pave the way for the development of practical Electronic Design Automation (EDA) agents, revolutionizing the field of board-level design.

cs.AR↗

arXiv2026

Low Rank Comes with Low Security: Gradient Assembly Poisoning Attacks against Distributed LoRA-based LLM Systems

Low-Rank Adaptation (LoRA) has become a popular solution for fine-tuning large language models (LLMs) in federated settings, dramatically reducing update costs by introducing trainable low-rank matrices. However, when integrated with frameworks like FedIT, LoRA introduces a critical vulnerability: clients submit $A$ and $B$ matrices separately, while only their product $AB$ determines the model update, yet this composite is never directly verified. We propose Gradient Assembly Poisoning (GAP), a novel attack that exploits this blind spot by crafting individually benign $A$ and $B$ matrices whose product yields malicious updates. GAP operates without access to training data or inter-client coordination and remains undetected by standard anomaly detectors. We identify four systemic vulnerabilities in LoRA-based federated systems and validate GAP across LLaMA, ChatGLM, and GPT-2. GAP consistently induces degraded or biased outputs while preserving surface fluency, reducing BLEU by up to 14.5\%, increasing factual and grammatical errors by over 800\%, and maintaining 92.6\% long-form response length. These results reveal a new class of stealthy, persistent threats in distributed LoRA fine-tuning.

cs.CR↗

arXiv2025

LIFT: Automating Symbolic Execution Optimization with Large Language Models for AI Networks

Dynamic Symbolic Execution (DSE) is a key technique in program analysis, widely used in software testing, vulnerability discovery, and formal verification. In distributed AI systems, DSE plays a crucial role in identifying hard-to-detect bugs, especially those arising from complex network communication patterns. However, traditional approaches to symbolic execution are often hindered by scalability issues and inefficiencies, particularly in large-scale systems. This paper introduces LIFT (Large-language-model Integrated Functional-equivalent-IR Transformation), a novel framework that leverages Large Language Models (LLMs) to automate the optimization of Intermediate Representations (IRs) in symbolic execution. LIFT addresses the challenges of symbolic execution by providing a scalable, context-sensitive solution for IR transformation. The framework consists of two phases: IR Analysis and Optimization, where LLMs optimize time-intensive IR blocks, and Symbolic Execution and Validation, which includes benchmarking and semantic verification to ensure correctness and generalizability. Experiments on real-world binaries demonstrated significant performance improvements, including a 53.5\% reduction in execution time for bigtest and a 10.24\% reduction for random, along with reductions in IR statements, PUT instructions, and temporary variables. These results demonstrate that LLMs simplify IRs while maintaining functional correctness, enhancing symbolic execution in distributed AI systems.

cs.CR↗

arXiv2024

VMID: A Multimodal Fusion LLM Framework for Detecting and Identifying Misinformation of Short Videos

Short video platforms have become important channels for news dissemination, offering a highly engaging and immediate way for users to access current events and share information. However, these platforms have also emerged as significant conduits for the rapid spread of misinformation, as fake news and rumors can leverage the visual appeal and wide reach of short videos to circulate extensively among audiences. Existing fake news detection methods mainly rely on single-modal information, such as text or images, or apply only basic fusion techniques, limiting their ability to handle the complex, multi-layered information inherent in short videos. To address these limitations, this paper presents a novel fake news detection method based on multimodal information, designed to identify misinformation through a multi-level analysis of video content. This approach effectively utilizes different modal representations to generate a unified textual description, which is then fed into a large language model for comprehensive evaluation. The proposed framework successfully integrates multimodal features within videos, significantly enhancing the accuracy and reliability of fake news detection. Experimental results demonstrate that the proposed approach outperforms existing models in terms of accuracy, robustness, and utilization of multimodal information, achieving an accuracy of 90.93%, which is significantly higher than the best baseline model (SV-FEND) at 81.05%. Furthermore, case studies provide additional evidence of the effectiveness of the approach in accurately distinguishing between fake news, debunking content, and real incidents, highlighting its reliability and robustness in real-world applications.

cs.CV↗

arXiv2024

A Depression Detection Method Based on Multi-Modal Feature Fusion Using Cross-Attention

Depression, a prevalent and serious mental health issue, affects approximately 3.8\% of the global population. Despite the existence of effective treatments, over 75\% of individuals in low- and middle-income countries remain untreated, partly due to the challenge in accurately diagnosing depression in its early stages. This paper introduces a novel method for detecting depression based on multi-modal feature fusion utilizing cross-attention. By employing MacBERT as a pre-training model to extract lexical features from text and incorporating an additional Transformer module to refine task-specific contextual understanding, the model's adaptability to the targeted task is enhanced. Diverging from previous practices of simply concatenating multimodal features, this approach leverages cross-attention for feature integration, significantly improving the accuracy in depression detection and enabling a more comprehensive and precise analysis of user emotions and behaviors. Furthermore, a Multi-Modal Feature Fusion Network based on Cross-Attention (MFFNC) is constructed, demonstrating exceptional performance in the task of depression identification. The experimental results indicate that our method achieves an accuracy of 0.9495 on the test dataset, marking a substantial improvement over existing approaches. Moreover, it outlines a promising methodology for other social media platforms and tasks involving multi-modal processing. Timely identification and intervention for individuals with depression are crucial for saving lives, highlighting the immense potential of technology in facilitating early intervention for mental health issues.

cs.CL↗

arXiv2024

Background-aware Multi-source Fusion Financial Trend Forecasting Mechanism

Stock prices, as an economic indicator, reflect changes in economic development and market conditions. Traditional stock price prediction models often only consider time-series data and are limited by the mechanisms of the models themselves. Some deep learning models have high computational costs, depend on a large amount of high-quality data, and have poor interpretations, making it difficult to intuitively understand the driving factors behind the predictions. Some studies have used deep learning models to extract text features and combine them with price data to make joint predictions, but there are issues with dealing with information noise, accurate extraction of text sentiment, and how to efficiently fuse text and numerical data. To address these issues in this paper, we propose a background-aware multi-source fusion financial trend forecasting mechanism. The system leverages a large language model to extract key information from policy and stock review texts, utilizing the MacBERT model to generate feature vectors. These vectors are then integrated with stock price data to form comprehensive feature representations. These integrated features are input into a neural network comprising various deep learning architectures. By integrating multiple data sources, the system offers a holistic view of market dynamics. It harnesses the comprehensive analytical and interpretative capabilities of large language models, retaining deep semantic and sentiment information from policy texts to provide richer input features for stock trend prediction. Additionally, we compare the accuracy of six models (LSTM, BiLSTM, MogrifierLSTM, GRU, ST-LSTM, SwinLSTM). The results demonstrate that our system achieves generally better accuracy in predicting stock movements, attributed to the incorporation of large language model processing, policy information, and other influential features.

cs.CE↗

arXiv2023

DefectHunter: A Novel LLM-Driven Boosted-Conformer-based Code Vulnerability Detection Mechanism

One of the most pressing threats to computing systems is software vulnerabilities, which can compromise both hardware and software components. Existing methods for vulnerability detection remain suboptimal. Traditional techniques are both time-consuming and labor-intensive, while machine-learning-based approaches often underperform when applied to complex datasets, due to their inability to capture high-dimensional relationships. Previous deep-learning strategies also fall short in capturing sufficient feature information. Although self-attention mechanisms can process information over long distances, they fail to capture structural information. In this paper, we introduce DefectHunter, an innovative model for vulnerability identification that employs the Conformer mechanism. This mechanism fuses self-attention with convolutional networks to capture both local, position-wise features and global, content-based interactions. Furthermore, we optimize the self-attention mechanisms to mitigate the issue of excessive attention heads introducing extraneous noise by adjusting the denominator. We evaluated DefectHunter against ten baseline methods using six industrial and two highly complex datasets. On the QEMU dataset, DefectHunter exhibited a 20.62\% improvement in accuracy over Pongo-70B, and for the CWE-754 dataset, its accuracy was 14.64\% higher. To investigate how DefectHunter comprehends vulnerabilities, we conducted a case study, which revealed that our model effectively understands the mechanisms underlying vulnerabilities.

cs.CR↗

arXiv2023

JFinder: A Novel Architecture for Java Vulnerability Identification Based Quad Self-Attention and Pre-training Mechanism

Software vulnerabilities pose significant risks to computer systems, impacting our daily lives, productivity, and even our health. Identifying and addressing security vulnerabilities in a timely manner is crucial to prevent hacking and data breaches. Unfortunately, current vulnerability identification methods, including classical and deep learning-based approaches, exhibit critical drawbacks that prevent them from meeting the demands of the contemporary software industry. To tackle these issues, we present JFinder, a novel architecture for Java vulnerability identification that leverages quad self-attention and pre-training mechanisms to combine structural information and semantic representations. Experimental results demonstrate that JFinder outperforms all baseline methods, achieving an accuracy of 0.97 on the CWE dataset and an F1 score of 0.84 on the PROMISE dataset. Furthermore, a case study reveals that JFinder can accurately identify four cases of vulnerabilities after patching.

cs.CR↗

arXiv2022

DeepVulSeeker: A Novel Vulnerability Identification Framework via Code Graph Structure and Pre-training Mechanism

Software vulnerabilities can pose severe harms to a computing system. They can lead to system crash, privacy leakage, or even physical damage. Correctly identifying vulnerabilities among enormous software codes in a timely manner is so far the essential prerequisite to patch them. Unfortantely, the current vulnerability identification methods, either the classic ones or the deep-learning-based ones, have several critical drawbacks, making them unable to meet the present-day demands put forward by the software industry. To overcome the drawbacks, in this paper, we propose DeepVulSeeker, a novel fully automated vulnerability identification framework, which leverages both code graph structures and the semantic features with the help of the recently advanced Graph Representation Self-Attention and pre-training mechanisms. Our experiments show that DeepVulSeeker not only reaches an accuracy as high as 0.99 on traditional CWE datasets, but also outperforms all other exisiting methods on two highly-complicated datasets. We also testified DeepVulSeeker based on three case studies, and found that DeepVulSeeker is able to understand the implications of the vulnerbilities. We have fully implemented DeepVulSeeker and open-sourced it for future follow-up research.

cs.CR↗

arXiv2022

Multi-View Pre-Trained Model for Code Vulnerability Identification

Vulnerability identification is crucial for cyber security in the software-related industry. Early identification methods require significant manual efforts in crafting features or annotating vulnerable code. Although the recent pre-trained models alleviate this issue, they overlook the multiple rich structural information contained in the code itself. In this paper, we propose a novel Multi-View Pre-Trained Model (MV-PTM) that encodes both sequential and multi-type structural information of the source code and uses contrastive learning to enhance code representations. The experiments conducted on two public datasets demonstrate the superiority of MV-PTM. In particular, MV-PTM improves GraphCodeBERT by 3.36\% on average in terms of F1 score.

cs.SE↗

arXiv2022

Solving the Federated Edge Learning Participation Dilemma: A Truthful and Correlated Perspective

An emerging computational paradigm, named federated edge learning (FEL), enables intelligent computing at the network edge with the feature of preserving data privacy for edge devices. Given their constrained resources, it becomes a great challenge to achieve high execution performance for FEL. Most of the state-of-the-arts concentrate on enhancing FEL from the perspective of system operation procedures, taking few precautions during the composition step of the FEL system. Though a few recent studies recognize the importance of FEL formation and propose server-centric device selection schemes, the impact of data sizes is largely overlooked. In this paper, we take advantage of game theory to depict the decision dilemma among edge devices regarding whether to participate in FEL or not given their heterogeneous sizes of local datasets. For realizing both the individual and global optimization, the server is employed to solve the participation dilemma, which requires accurate information collection for devices' local datasets. Hence, we utilize mechanism design to enable truthful information solicitation. With the help of correlated equilibrium, we derive a decision making strategy for devices from the global perspective, which can achieve the long-term stability and efficacy of FEL. For scalability consideration, we optimize the computational complexity of the basic solution to the polynomial level. Lastly, extensive experiments based on both real and synthetic data are conducted to evaluate our proposed mechanisms, with experimental results demonstrating the performance advantages.

cs.NI↗

arXiv2020

Tokoin: A Coin-Based Accountable Access Control Scheme for Internet of Things

With the prevalence of Internet of Things (IoT) applications, IoT devices interact closely with our surrounding environments, bringing us unparalleled smartness and convenience. However, the development of secure IoT solutions is getting a long way lagged behind, making us exposed to common unauthorized accesses that may bring malicious attacks and unprecedented danger to our daily life. Overprivilege attack, a widely reported phenomenon in IoT that accesses unauthorized or excessive resources, is notoriously hard to prevent, trace and mitigate. To tackle this challenge, we propose Tokoin-Based Access Control (TBAC), an accountable access control model enabled by blockchain and Trusted Execution Environment (TEE) technologies, to offer fine-graininess, strong auditability, and access procedure control for IoT. TBAC materializes the virtual access power into a definite-amount and secure cryptographic coin termed "tokoin" (token+coin), and manages it using atomic and accountable state-transition functions in a blockchain. We also realize access procedure control by mandating every tokoin a fine-grained access policy defining who is allowed to do what at when in where by how. The tokoin is peer-to-peer transferable, and can be modified only by the resource owner when necessary. We fully implement TBAC with well-studied cryptographic primitives and blockchain platforms and present a readily available APP for regular users. We also present a case study to demonstrate how TBAC is employed to enable autonomous in-home cargo delivery while guaranteeing the access policy compliance and home owner's physical security by regulating the physical behaviors of the deliveryman.

cs.CR↗