Selected work

Representative Papers

MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection

Aug 02, 2026

This work addresses the challenge of multimodal fake news detection for emerging events in zero-shot scenarios, where existing methods often overlook the reuse patterns of historical disinformation and subtle cross-modal inconsistencies. To tackle this, we propose the first framework that integrates retrieval augmentation with multi-agent collaborative debate: it first retrieves relevant historical news articles via multimodal similarity search, then leverages bidirectional evidence reasoning within a structured debate mechanism to assess veracity. By effectively capturing cross-modal discrepancies and characteristic patterns of past deceptive strategies, our approach significantly outperforms state-of-the-art methods across three benchmark datasets, achieving an accuracy improvement of up to 2.35% on Weibo-21.

0 citationsRead paper

SCARA: A Semantics-Constrained Autonomous Remediation Agent for Opaque Industrial Software Vulnerabilities

May 19, 2026

This work proposes SCARA, the first end-to-end autonomous repair agent designed to address the challenge of automatically patching vulnerabilities in opaque industrial software (OIS)—systems that lack source code, symbols, and recompilability. Operating entirely at the binary level, SCARA employs a three-stage mechanism comprising operational state-aware validation (OSVA), repair synthesis under semantic constraints (RSA), and correctness verification (CVA). It integrates protocol-level mitigation, binary hardening, and SSCKG-guided patch generation to ensure both feasibility and semantic correctness of repairs. Evaluated on the OIS-RemedBench benchmark, SCARA achieves 100% repair precision with zero false positives, filters out 20.0% of infeasible cases, and attains an 88.9% final repair success rate after retries.

0 citationsRead paper

Securing the Dark Matter: A Semantic-Enhanced Neuro-Symbolic Framework for Supply Chain Analysis of Opaque Industrial Software

May 08, 2026

Industrial critical infrastructure software is often deployed as stripped, unsigned binaries, rendering traditional software composition analysis ineffective and challenging existing methods to jointly capture structural and behavioral semantics. This work proposes a semantics-enhanced neuro-symbolic framework that integrates abstract interpretation with a reflective prompting mechanism to constrain a local large language model and suppress hallucinations. It employs a surjective transformation to compress code property graphs into typed software supply chain knowledge graphs and introduces a domain-adapted Graphormer with embedding-space subgraph matching to enable global risk reasoning and zero-day/APT attack identification in opaque binaries. Evaluated on three progressively challenging benchmarks and a hybrid physical-virtual platform comprising real devices from five industrial control vendors, the approach significantly improves detection rates for high-severity CVEs and semantic fidelity while substantially reducing false positives.

0 citationsRead paper

OmniFM: Toward Modality-Robust and Task-Agnostic Federated Learning for Heterogeneous Medical Imaging

Mar 23, 2026

This work addresses the challenges of substantial inter-institutional modality distribution shifts and diverse downstream tasks in medical image analysis under federated learning settings. To this end, we propose OmniFM, the first modality-robust and task-agnostic federated learning framework. OmniFM leverages low-frequency components in the frequency domain to model cross-modal consistency and integrates global spectral knowledge retrieval, embedding-level cross-attention fusion, and a pre-/post-spectral prompt mechanism to uniformly support multiple tasks—including classification, segmentation, super-resolution, and visual question answering—without modifying the training pipeline. The framework jointly models spectral priors and personalized prompts, further enhanced by spectral proximal alignment regularization. Extensive experiments on real-world medical datasets demonstrate that OmniFM consistently outperforms existing methods in both intra- and cross-modality heterogeneous scenarios, under both from-scratch training and fine-tuning setups.

0 citationsRead paper

MALF: A Multi-Agent LLM Framework for Intelligent Fuzzing of Industrial Control Protocols

Oct 02, 2025

Addressing the challenges of low detection efficacy for Industrial Control Protocol (ICP) vulnerabilities, poor coverage, and high false-positive rates in conventional fuzzing, this paper proposes MALF—a Multi-Agent Large-model Fuzzing framework. MALF introduces a novel multi-agent collaborative architecture that synergistically integrates domain-knowledge enhancement with protocol-aware generation. It incorporates Retrieval-Augmented Generation (RAG), QLoRA-based lightweight fine-tuning, feedback-driven mutation, and Shannon-entropy-guided diversity control to establish a dynamic closed loop spanning protocol understanding, seed generation, and feedback-based optimization. Evaluated on mainstream protocols including Modbus/TCP, MALF achieves 88–92% test pass rates and >90% seed coverage, uncovering three CNVD-certified zero-day vulnerabilities. Its vulnerability detection rate significantly outperforms state-of-the-art approaches.

0 citationsRead paper

Recent publications

Latest Papers

MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection

Aug 02, 2026

This work addresses the challenge of multimodal fake news detection for emerging events in zero-shot scenarios, where existing methods often overlook the reuse patterns of historical disinformation and subtle cross-modal inconsistencies. To tackle this, we propose the first framework that integrates retrieval augmentation with multi-agent collaborative debate: it first retrieves relevant historical news articles via multimodal similarity search, then leverages bidirectional evidence reasoning within a structured debate mechanism to assess veracity. By effectively capturing cross-modal discrepancies and characteristic patterns of past deceptive strategies, our approach significantly outperforms state-of-the-art methods across three benchmark datasets, achieving an accuracy improvement of up to 2.35% on Weibo-21.

0 citationsRead paper

SCARA: A Semantics-Constrained Autonomous Remediation Agent for Opaque Industrial Software Vulnerabilities

May 19, 2026

This work proposes SCARA, the first end-to-end autonomous repair agent designed to address the challenge of automatically patching vulnerabilities in opaque industrial software (OIS)—systems that lack source code, symbols, and recompilability. Operating entirely at the binary level, SCARA employs a three-stage mechanism comprising operational state-aware validation (OSVA), repair synthesis under semantic constraints (RSA), and correctness verification (CVA). It integrates protocol-level mitigation, binary hardening, and SSCKG-guided patch generation to ensure both feasibility and semantic correctness of repairs. Evaluated on the OIS-RemedBench benchmark, SCARA achieves 100% repair precision with zero false positives, filters out 20.0% of infeasible cases, and attains an 88.9% final repair success rate after retries.

0 citationsRead paper

Securing the Dark Matter: A Semantic-Enhanced Neuro-Symbolic Framework for Supply Chain Analysis of Opaque Industrial Software

May 08, 2026

Industrial critical infrastructure software is often deployed as stripped, unsigned binaries, rendering traditional software composition analysis ineffective and challenging existing methods to jointly capture structural and behavioral semantics. This work proposes a semantics-enhanced neuro-symbolic framework that integrates abstract interpretation with a reflective prompting mechanism to constrain a local large language model and suppress hallucinations. It employs a surjective transformation to compress code property graphs into typed software supply chain knowledge graphs and introduces a domain-adapted Graphormer with embedding-space subgraph matching to enable global risk reasoning and zero-day/APT attack identification in opaque binaries. Evaluated on three progressively challenging benchmarks and a hybrid physical-virtual platform comprising real devices from five industrial control vendors, the approach significantly improves detection rates for high-severity CVEs and semantic fidelity while substantially reducing false positives.

0 citationsRead paper

OmniFM: Toward Modality-Robust and Task-Agnostic Federated Learning for Heterogeneous Medical Imaging

Mar 23, 2026

This work addresses the challenges of substantial inter-institutional modality distribution shifts and diverse downstream tasks in medical image analysis under federated learning settings. To this end, we propose OmniFM, the first modality-robust and task-agnostic federated learning framework. OmniFM leverages low-frequency components in the frequency domain to model cross-modal consistency and integrates global spectral knowledge retrieval, embedding-level cross-attention fusion, and a pre-/post-spectral prompt mechanism to uniformly support multiple tasks—including classification, segmentation, super-resolution, and visual question answering—without modifying the training pipeline. The framework jointly models spectral priors and personalized prompts, further enhanced by spectral proximal alignment regularization. Extensive experiments on real-world medical datasets demonstrate that OmniFM consistently outperforms existing methods in both intra- and cross-modality heterogeneous scenarios, under both from-scratch training and fine-tuning setups.

0 citationsRead paper

MALF: A Multi-Agent LLM Framework for Intelligent Fuzzing of Industrial Control Protocols

Oct 02, 2025

Addressing the challenges of low detection efficacy for Industrial Control Protocol (ICP) vulnerabilities, poor coverage, and high false-positive rates in conventional fuzzing, this paper proposes MALF—a Multi-Agent Large-model Fuzzing framework. MALF introduces a novel multi-agent collaborative architecture that synergistically integrates domain-knowledge enhancement with protocol-aware generation. It incorporates Retrieval-Augmented Generation (RAG), QLoRA-based lightweight fine-tuning, feedback-driven mutation, and Shannon-entropy-guided diversity control to establish a dynamic closed loop spanning protocol understanding, seed generation, and feedback-based optimization. Evaluated on mainstream protocols including Modbus/TCP, MALF achieves 88–92% test pass rates and >90% seed coverage, uncovering three CNVD-certified zero-day vulnerabilities. Its vulnerability detection rate significantly outperforms state-of-the-art approaches.

0 citationsRead paper