Binary analysis increasingly relies on large language models (LLMs) to perform semantic reasoning over complex program behaviors. However, existing approaches largely adopt a one-pass execution paradigm, where reasoning operates over a fixed program representation constructed by static analysis tools. This formulation limits the ability to adapt exploration based on intermediate results and makes it difficult to sustain long-horizon, multi-path analysis under constrained context. We present FORGE, a system that rethinks LLM-based analysis as a feedback-driven execution process. FORGE interleaves reasoning and tool interaction through a reasoning-action-observation loop, enabling incremental exploration and evidence construction. To address the instability of long-horizon reasoning, we introduce a Dynamic Forest of Agents (FoA), a decomposed execution model that dynamically coordinates parallel exploration while bounding per-agent context. We evaluate FORGE on 3,457 real-world firmware binaries. FORGE identifies 1,274 vulnerabilities across 591 unique binaries, achieving 72.3

Binary vulnerability analysis is increasingly performed by LLM-based agents in an iterative, multi-pass manner, with the model as the core decision-maker. However, how such systems organize exploration over hundreds of reasoning steps remains poorly understood, due to limited context windows and implicit token-level behaviors. We present the first large-scale, trace-level study showing that multi-pass LLM reasoning gives rise to structured, token-level implicit patterns. Analyzing 521 binaries with 99,563 reasoning steps, we identify four dominant patterns: early pruning, path-dependent lock-in, targeted backtracking, and knowledge-guided prioritization that emerge implicitly from reasoning traces. These token-level implicit patterns serve as an abstraction of LLM reasoning: instead of explicit control-flow or predefined heuristics, exploration is organized through implicit decisions regulating path selection, commitment, and revision. Our analysis shows these patterns form a stable, structured system with distinct temporal roles and measurable characteristics. Our results provide the first systematic characterization of LLM-driven binary analysis and a foundation for more reliable analysis systems.

Unstructured pruning introduces significant sparsity in deep neural networks (DNNs), enhancing accelerator hardware efficiency. However, three critical challenges constrain performance gains: 1) complex fetching logic for nonzero (NZ) data pairs; 2) load imbalance across processing elements (PEs); and 3) PE stalls from write-back contention. This brief proposes an energy-efficient accelerator addressing these inefficiencies through three innovations. First, we propose a Cartesian-product output-row-stationary (CPORS) dataflow that inherently matches NZ data pairs by sequentially fetching compressed data. Second, a multilevel partial sum reduction (MLPR) strategy minimizes write-back traffic and converts random PE stalls into manageable load imbalance. Third, a kernel sorting and load scheduling (KSLS) mechanism resolves PE idle/stall and achieves PE array-level load balancing, attaining 76.6% average PE utilization across all sparsity levels. Implemented in 22-nmCMOS, the accelerator delivers 1.85x speedup and 1.4x energy efficiency over baseline and achieves 25.8 TOPS/W peak energy efficiency at 90% sparsity.

The time-delay attacks pose serious security threats to the industrial control systems (ICSs), where ICS infrastructures (e.g., chemical factories) could suffer severe safety consequences. They could bypass current delay detection methods by avoiding triggering packet timeouts. In this article, we reveal that malicious states caused by the time-delay attacks in ICS scenarios can be detected by analyzing ICS programs. We propose detecting a time-delay attack in ICS scenarios by comparing the difference between malicious and benign states, meeting the real-time and noninterference requirements. Specifically, we utilize symbolic execution to analyze ICS programs to generate the benign states of ICS and leverage the key features of time-delay attacks to create the malicious states of ICS, where the states are transferred through the network for remote control and monitoring. We propose a multimodal neural network whose inputs are the malicious states sampled from the ICS network traffic and the time domain features, and the output is whether such a time-delay attack exists. We implement a prototype system and conduct real-world experiments to evaluate the performance of our detection approach. Our experiments cover 102 vulnerable ICS programs and five types of time-delay attacks. The evaluation results show that our approach can detect ICS time-delay attacks in 0.6 s, with 97.2% precision and 98% recall.

Today’s security tools predominantly rely on predefined rules crafted by experts, making them poorly adapted to the emergence of software supply chain attacks. To tackle this limitation, we propose a novel tool, RuleLLM, which leverages large language models (LLMs) to automate rule generation for OSS ecosystems. RuleLLM extracts metadata and code snippets from malware as its input, producing YARA and Semgrep rules that can be directly deployed in software development. Specifically, the rule generation task involves three subtasks: crafting rules, refining rules, and aligning rules. To validate RuleLLM’s effectiveness, we implemented a prototype system and conducted experiments on the dataset of 1,633 malicious packages. The results are promising—RuleLLM generated 763 rules (452 YARA and 311 Semgrep) with a precision of 85.2% and a recall of 91.8%, outperforming state-of-the-art (SOTA) tools and scored-based approaches. We further analyzed generated rules and proposed a rule taxonomy: 11 categories and 38 subcategories.

Large Language Models (LLMs) and their agent systems have recently demonstrated strong potential in automating code reasoning and vulnerability detection. However, when applied to large-scale firmware, their performance degrades due to the binary nature of firmware, complex dependency structures, and heterogeneous components. To address this challenge, this paper presents FIRMHIVE, a recursive agent hive that enables LLMs to act as autonomous firmware security analysts. FIRMHIVE introduces two key mechanisms: (1) transforming delegation into a per-agent, executable primitive and (2) constructing a runtime Tree of Agents (ToA) for decentralized coordination. We evaluate FIRMHIVE using real-world firmware images obtained from publicly available datasets, covering five representative security analysis tasks. Compared with existing LLM-agent baselines, FIRMHIVE performs deeper (about 16x more reasoning steps) and broader (about 2.3x more files inspected) cross-file exploration, resulting in about 5.6x more alerts per firmware. Compared to state-of-the-art (SOTA) security tools, FIRMHIVE identifies about 1.5x more vulnerabilities (1,802 total) and achieves 71

Telephone scams, with their profound psychological impact, often compel victims to make hasty and severe decisions. Studying these scams is challenging due to the scarcity of comprehensive datasets, a result of the private nature of telephone interactions. In this paper, we introduce ScamGen, a template-based data augmentation technique designed to enhance Chinese telephone scam data. ScamGen leverages psychological insights to generate diverse and realistic scam scenarios, focusing on the psychological dynamics between scammers and victims. This novel approach integrates psychological theory with data augmentation, diverging from traditional methods by emphasizing scammer-victim interactions. Our method begins with a multi-source data collection framework, compiling an initial seed dataset of tele-scam samples. Using sentence- and word-level perturbations, we expand this seed data to create a comprehensive and diverse dataset covering a wide range of scam scenarios. Rigorous evaluations demonstrate that ScamGen outperforms large language models in generating high-quality, varied datasets. Additionally, we develop five deep learning models for intent detection on this dataset, with BERT achieving the highest precision at 86.68%. The dataset, which will be made publicly available, marks a significant step toward understanding scammer tactics and improving tele-scam detection systems.

Fingerprinting is a network reconnaissance technique utilized for gathering information about online computing systems, including operation systems and applications. Unfortunately, attackers typically leverage fingerprinting techniques to locate, enumerate, and subsequently target vulnerable systems, which is the first primary stage of a cyber attack. In this work, we explore the susceptibility of machine learning (ML)-based classifiers to misclassification, where a slight perturbation in the packet is included to spoof OS fingerprints. We propose SOFI (Spoof OS Fingerprints), an adversarial example generation algorithm under TCP/IP specification constraints, to create effective perturbations in a packet for deceiving an OS fingerprint. Specifically, SOFI has three major technical innovations: (1) it is the first to utilize adversarial examples to automatically perturb fingerprinting techniques; (2) it complies with constraints and integrity of network packets; (3) it achieves a high success rate in spoofing OS fingerprints. We validate the effectiveness of adversarial packets against active and passive OS fingerprints, verifying the transferability and robustness of SOFI. Comprehensive experimental results demonstrate that SOFI automatically identifies applicable and available OS fingerprint features, unlike existing tools relying on expert knowledge.

The existing limitations of the fundamental laws necessary for constructing a comprehensive and widely accepted theoretical framework have significantly hindered the progress of Information Management. This lack has resulted in a predominant reliance on indirect strategies to address information management challenges, often leading to complex, inefficient, and somewhat stochastic analyses and evaluations. For instance, the failure rate of digital transformation in global enterprises is as high as 80 %, and that of data-driven organizational change reaches 85 %, highlighting the urgency and difficulty of resolving these challenges. Through an in-depth analysis of the spiral model and derivation of the Shannon-Weaver model, we unearthed the objective and universal Compendium Law of iterative information management. Building on this law, we propose the application of information system modeling and Hamiltonian graph theory to develop a comprehensive analytical model for iterative information management. This model provides a theoretical approach for the scientific analysis and optimal design of iterative information management, enabling efficient comparative analysis and knowledge transfer among various iterative information management systems. This study contributes to the foundational understanding of Information Management as an independent discipline capable of addressing cross-disciplinary challenges related to information resources, including those found in artificial intelligence, blockchains, quantum communication, the Internet of Things, and digitization.

The open-source software (OSS) ecosystem suffers from security threats caused by malware. However, OSS malware research has three limitations: a lack of high-quality datasets, a lack of malware diversity, and a lack of attack campaign contexts. In this paper, we first build the largest dataset of 24,356 malicious packages from online sources, then propose a knowledge graph to represent the OSS malware corpus and conduct malware analysis in the wild. Our main findings include (1) it is essential to collect malicious packages from various online sources because their data overlapping degrees are small; (2) despite the sheer volume of malicious packages, many reuse similar code, leading to a low diversity of malware; (3) only 28 malicious packages were repeatedly hidden via dependency libraries of 1,354 malicious packages, and dependency-hidden malware has a shorter active time; (4) security reports are the only reliable source for disclosing the malware-based context.

Fault diagnosis of mechanical equipment plays a critical role in enhancing system stability and ensuring operational safety. Multi-modal monitoring data provides a more comprehensive view of the equipment's condition, enabling more precise state diagnosis through the processing of such data. A major challenge in contemporary multi-modal information fusion lies in the effective integration of cross-modal information. To overcome the inherent limitations of single-view approaches often found in existing multi-modal methods, this paper presents a multi-view, multi-modal information fusion approach based on graph convolutional networks. This method not only extracts key features from signals efficiently but also uncovers the interrelationships between different modal signals through a multi-view interaction mechanism, achieving more robust information fusion and enhanced fault diagnosis performance. This method consists of three core modules: feature extraction, information fusion, and classification. The feature extraction module utilizes a multi-level residual architecture, Mini-Long Short Term Memory, and self-attention mechanisms to capture both local and global signal features, model temporal dependencies, and refine feature selection. The information fusion module combines data and feature-level signal correlations using graph convolutional networks for cross-modal interaction. The classification module employs a two-layer fully connected network for fault diagnosis. This method quantitatively assesses the contribution of each modality, providing a theoretical basis for understanding their physical significance. Experimental results and ablation studies demonstrate its superior performance and enhanced accuracy in fault diagnosis, offering a novel approach for condition monitoring of complex equipment.

While vulnerability databases are important sources of information for software security, it is known that information in these databases is inconsistent. How to rectify these incorrect data is a challenging issue. In this article, we employ developer logs and patches to automatically identify vulnerable source code versions that each CVE really affects. Our tool organizes all versions of a piece of software into a version tree, and identifies the first vulnerable version, and the last vulnerable versions in the version tree trunk and branches. For evaluation, we took Linux Kernel as the case study and quantified the error rate of the vulnerable versions reported by the NVD. The total number of vulnerable Linux Kernel versions reported by the NVD was 43,727 (as of September 2020), of which the total number of false positives reached 2,497 and the total number of false negatives reached 9,330, accounting for 5.7% and 21.34%, respectively. In addition, we compare our tool with two vulnerability detection tools and show that our tool could achieve high detection accuracy.

The advent of Machine Learning as a Service (MLaaS) and deep learning applications has increased the susceptibility of models to adversarial textual attacks, particularly in black -box settings. Prior work on black -box adversarial textual attacks generally follows a stable strategy that involves leveraging char-level, world-level, and sentence-level perturbations, as well as using queries to the target model to find adversarial examples in the search space. However, existing approaches prioritize query efficiency by reducing the search space, thereby overlooking hard-to-attack textual instances. To address this issue, we propose BFS2Adv , a brute force algorithm that generates adversarial examples for both easy-to-attack and hard-to-attack textual inputs. BFS2Adv, starting with an original text, employs word-level perturbations and synonym substitution to construct a comprehensive search space, with each node representing a potential adversarial example. The algorithm systematically explores this space through a breadth-first search, combined with queries to the target model, to effectively identify qualified adversarial examples. We implemented and evaluated a prototype of BFS2Adv against renowned models such as ALBERT and BERT, utilizing the SNLI and MR datasets. Our results demonstrate that BFS2Adv outperforms state -of -the -art algorithms and effectively improves the success rate of short -text adversarial attacks. Furthermore, we provide detailed insights into the robustness of BFS2Adv by analyzing those hard-to-attack examples.

The Internet of Things (IoT) has become indispensable to our daily lives and work. Unfortunately, developers often reuse software libraries in the IoT firmware, leading to a major security concern. If vulnerabilities or insecure versions of these libraries go unpatched, a massive number of IoT devices can be impacted. In this paper, we propose the AutoFirm, an automated tool for detecting reused libraries in IoT firmware at a large scale. Specifically, AutoFirm leverages the syntax information (library name and version) to determine whether IoT firmware reuses the libraries. We conduct a large-scale empirical study of reused libraries of IoT firmware, investigating more than 6,900+ firmware and 2,700+ distinct vulnerabilities affecting 11,300+ vulnerable versions from 349 open-source software libraries. Leveraging this diverse information set, we conduct a qualitative assessment of vulnerable library versions to understand security gaps and the misplaced trust of libraries in IoT firmware. Our research reveals that: manufacturers neglected to update outdated libraries for IoT firmware in 67.3\% of cases; on average, outdated libraries persisted for over 1.34 years prior to remediation; vulnerabilities of software libraries have posed server threats to widespread IoT devices.

Nowadays, the open-source software (OSS) ecosystem suffers from security threats of software supply chain (SSC) attacks. Interpreted OSS malware plays a vital role in SSC attacks, as criminals have an arsenal of attack vectors to deceive users into installing malware and executing malicious activities. In this paper, we introduce tactics, techniques, and procedures (TTPs) proposed by MITRE ATT&CK into the interpreted malware analysis to characterize different phases of an attack lifecycle. Specifically, we propose GENTTP, a zero-shot approach to extracting a TTP of an interpreted malware package. GENTTP leverages large language models (LLMs) to automatically generate a TTP, where the input is a malicious package, and the output is a deceptive tactic and an execution tactic of attack vectors. To validate the effectiveness of GENTTP, we collect two datasets for evaluation: a dataset with ground truth labels and a large dataset in the wild. Experimental results show that GENTTP can generate TTPs with high accuracy and efficiency. To demonstrate GENTTP's benefits, we build an LLM-based Chatbot from 3,700+ PyPI malware's TTPs. We further conduct a quantitative analysis of malware's TTPs at a large scale. Our main findings include: (1) many OSS malicious packages share a relatively stable TTP, even with the increasing emergence of malware and attack campaigns, (2) a TTP reflects characteristics of a malware-based attack, and (3) an attacker's intent behind the malware is linked to a TTP.

Package managers such as NPM, Maven, and PyPI play a pivotal role in open-source software (OSS) ecosystems, streamlining the distribution and management of various freely available packages. The fine-grained details within software packages can unveil potential risks within existing OSS ecosystems, offering valuable insights for detecting malicious packages. In this study, we undertake a large-scale empirical analysis focusing on fine-grained information (FGI): the metadata, static, and dynamic functions. Specifically, we investigate the FGI usage across a diverse set of 50,000+ legitimate and 1,000+ malicious packages. Based on this diverse data collection, we conducted a comparative analysis between legitimate and malicious packages. Our findings reveal that (1) malicious packages have less metadata content and utilize fewer static and dynamic functions than legitimate ones; (2) malicious packages demonstrate a higher tendency to invoke HTTP/URL functions as opposed to other application services, such as FTP or SMTP; (3) FGI serves as a distinguishable indicator between legitimate and malicious packages; and (4) one dimension in FGI has sufficient distinguishable capability to detect malicious packages, and combining all dimensions in FGI cannot significantly improve overall performance.

Nowadays security vulnerability reports contain commercial vendor-centric information but fail to include accurate information of open-source software packages. Open-source ecosystems use package managers, such as Maven, NuGet, NPM, and Gem, to cover hundreds of thousands of free code packages. However, we uncover that vulnerability reports frequently miss the vulnerable software package information when the software package comes from open-source ecosystems. To fill in this gap, we propose a framework called PKVIC (software p ac k age v ulnerability i nformation c alibration), as the first tool to automatically associate security vulnerability reports with affected software packages from different open-source ecosystems. Specifically, PKVIC designs an ecosystem classifier to determine which ecosystem a vulnerability report belongs to. From the reports written in natural language, PKVIC extracts the entities closely related to software names in ecosystems. To efficiently and accurately locate the affected software packages from millions of packages, we propose a recursive traversal method to generate the package identifier based on the naming scheme and candidate named entities. We implemented the prototype of PKVIC and conducted comprehensive experiments to validate its efficacy. In particular, we ran PKVIC over 421,808 vulnerability reports from 20 well-known sources of security vulnerabilities and identified 11,279 unique vulnerability reports that affected 2,703 open-source software packages. PKVIC successfully found the accurate reference URLs for these 2,703 software packages across 6 open-source ecosystems, including Pypi, Gem, NPM, Packagist, Nuget, and Maven.

Closed-circuit television (CCTV) systems and surveillance devices have been widely used for monitoring the real world (e.g., streets, lanes, buildings). Illumination of the nature of surveillance devices’ images can help in deterring criminal activity and provide insights for improving the infrastructure safety of CCTV systems. However, millions of outdated surveillance devices are still in use today, which suffer noise, the field of view, low-light conditions, and low visual quality. In this work, we conduct a large-scale empirical study of images of surveillance devices in the wild. For data collection, we deploy a web crawler to gather webcam images in the wild covering 100 websites that distribute the live streams to the public, such as a plaza and an industrial area. To enhance the low-light quality of images, we propose a new framework called ImCam for enhancing the quality of webcam images in the wild, including the retinex model and generative adversarial network (GAN). We develop a prototype of ImCam and evaluate its effectiveness over three classification systems, AlexNet, ResNet, and Vision Transformer (ViT), and Machine Learning as a Service (MLaaS) in three cloud service providers (Azure, Baidu, and Tencent). Our results show that ImCam is capable of improving low-light images for CCTV systems, making them acceptable in practice. Finally, leveraging 203,786 live streams and our ImCam, we conduct an analysis of webcam images in the wild and shed light on the nature of security usage for CCTV systems.

Download This Paper Open PDF in Browser Add Paper to My Library Share: Permalink Using these links will ensure access to this page indefinitely Copy URL Copy DOI