Institution profile

Agency for Defense Development

Academic institutionasia · kr
Official website
Research library24linked papers
Opportunities0open roles
Selected work

Representative Papers

What Does Fr\'echet Distance Measure? A Directional Decomposition

Oct 04, 2026

This work addresses the limitation of the traditional Fréchet distance, which evaluates generative models via a single scalar that obscures the specific causes of distributional discrepancies and disconnects the metric from perceptual quality. We introduce the Directional Fréchet Distance, which decomposes optimal transport displacements through projection and leverages multimodal embeddings such as CLIP to map abstract distances onto interpretable semantic directions, revealing that a few key dimensions dominate most deviations. This approach successfully resolves conflicts between FID and human preferences in diffusion models, quantifies frame-wise appearance discrepancies in video FVD, and reinterprets the physical meaning of protein FID. By enabling fine-grained attribution analysis of generative quality across domains, this method provides a principled diagnostic tool for generative modeling. The code is publicly available.

0 citationsRead paper

Diffusion ReRoll: Revisable Denoising for Robotic Sequential Prediction

Jul 22, 2026

Existing diffusion-based approaches for robotic sequential prediction rely on unidirectional denoising, which struggles to maintain global consistency in long-horizon tasks. This work proposes a reversible denoising mechanism that incorporates a structured re-noising strategy, selectively re-adding noise to temporally stable local regions during the diffusion process. This enables iterative refinement by integrating cross-temporal contextual information, facilitating mutual correction between early and late segments of the predicted sequence. Implemented within a unified video-action diffusion framework, the method combines context-aware denoising with selective re-noising, achieving up to a 56.5% improvement in average success rate on the OGBench and LIBERO-10 benchmarks. Moreover, it demonstrates enhanced robustness and stronger action-video consistency in out-of-distribution scenarios.

0 citationsRead paper

GARAGE: Characterizing the Automation Boundary in LLM-based Attack Graph Generation

Jul 20, 2026

Existing automated tools struggle to effectively process unstructured cyber threat intelligence (CTI) and the unique architectures of automotive systems, hindering vehicle-specific attack graph generation. This work proposes GARAGE, a novel framework that integrates retrieval-augmented generation (RAG) with domain-specific automotive security knowledge. Built upon 12,786 CVE entries and 140 incident reports, GARAGE constructs a knowledge base compliant with STIX 2.1 and Auto-ISAC ATM standards, enabling fine-grained kill-chain analysis for tactical-level attack scenario modeling. The approach supports generalization to unseen vehicle architectures and demonstrates accurate knowledge transfer across 320 leave-one-out experiments. Furthermore, it delineates the capability boundaries of large language models in threat analysis and provides cost-performance deployment strategies, thereby enhancing human-machine collaborative TARA processes.

0 citationsRead paper
Recent publications

Latest Papers

What Does Fr\'echet Distance Measure? A Directional Decomposition

Oct 04, 2026

This work addresses the limitation of the traditional Fréchet distance, which evaluates generative models via a single scalar that obscures the specific causes of distributional discrepancies and disconnects the metric from perceptual quality. We introduce the Directional Fréchet Distance, which decomposes optimal transport displacements through projection and leverages multimodal embeddings such as CLIP to map abstract distances onto interpretable semantic directions, revealing that a few key dimensions dominate most deviations. This approach successfully resolves conflicts between FID and human preferences in diffusion models, quantifies frame-wise appearance discrepancies in video FVD, and reinterprets the physical meaning of protein FID. By enabling fine-grained attribution analysis of generative quality across domains, this method provides a principled diagnostic tool for generative modeling. The code is publicly available.

0 citationsRead paper

Diffusion ReRoll: Revisable Denoising for Robotic Sequential Prediction

Jul 22, 2026

Existing diffusion-based approaches for robotic sequential prediction rely on unidirectional denoising, which struggles to maintain global consistency in long-horizon tasks. This work proposes a reversible denoising mechanism that incorporates a structured re-noising strategy, selectively re-adding noise to temporally stable local regions during the diffusion process. This enables iterative refinement by integrating cross-temporal contextual information, facilitating mutual correction between early and late segments of the predicted sequence. Implemented within a unified video-action diffusion framework, the method combines context-aware denoising with selective re-noising, achieving up to a 56.5% improvement in average success rate on the OGBench and LIBERO-10 benchmarks. Moreover, it demonstrates enhanced robustness and stronger action-video consistency in out-of-distribution scenarios.

0 citationsRead paper

GARAGE: Characterizing the Automation Boundary in LLM-based Attack Graph Generation

Jul 20, 2026

Existing automated tools struggle to effectively process unstructured cyber threat intelligence (CTI) and the unique architectures of automotive systems, hindering vehicle-specific attack graph generation. This work proposes GARAGE, a novel framework that integrates retrieval-augmented generation (RAG) with domain-specific automotive security knowledge. Built upon 12,786 CVE entries and 140 incident reports, GARAGE constructs a knowledge base compliant with STIX 2.1 and Auto-ISAC ATM standards, enabling fine-grained kill-chain analysis for tactical-level attack scenario modeling. The approach supports generalization to unseen vehicle architectures and demonstrates accurate knowledge transfer across 320 leave-one-out experiments. Furthermore, it delineates the capability boundaries of large language models in threat analysis and provides cost-performance deployment strategies, thereby enhancing human-machine collaborative TARA processes.

0 citationsRead paper