Selected work

Representative Papers

TooBad: Backdoor Diffusion Models with Ultra-Low Poison Rate and Imperceptible Trigger

Jun 22, 2026

Existing backdoor attacks on diffusion models struggle to simultaneously achieve high efficiency, low poisoning rates, and strong stealthiness. This work proposes TooBad, a novel framework that introduces, for the first time, a trigger optimization mechanism tailored specifically for diffusion models. By integrating fine-tuning with minimal poisoned data injection, TooBad attains over 85% attack success rate with only a 0.5% poisoning ratio; when the poisoning ratio increases to 5%, near-perfect success (≈100%) is achieved within just 3–5 training epochs. The method substantially reduces training overhead while effectively evading state-of-the-art defense mechanisms, thereby offering both potent attack performance and high concealment.

0 citationsRead paper

Critical-CoT: A Robust Defense Framework against Reasoning-Level Backdoor Attacks in Large Language Models

Apr 12, 2026

Large language models are vulnerable to reasoning-level backdoor attacks, wherein adversaries embed malicious reasoning steps into the chain-of-thought via trigger mechanisms, leading models to produce seemingly plausible yet harmful outputs. This work proposes Critical-CoT, a novel defense framework specifically designed to counter such attacks. Critical-CoT employs a two-stage fine-tuning strategy to endow models with critical thinking capabilities, enabling them to automatically detect and reject compromised reasoning steps. Experimental results demonstrate that the proposed method achieves strong robustness against both in-context learning and fine-tuning-based backdoor attacks across multiple mainstream large language models and datasets. Furthermore, it significantly enhances model safety while exhibiting excellent generalization across domains and tasks.

0 citationsRead paper

Agentic AI Meets Edge Computing in Autonomous UAV Swarms

Jan 20, 2026IEEE Internet of Things Magazine

This work addresses the challenge of achieving efficient and autonomous coordination among drone swarms in high-risk, infrastructure-constrained dynamic environments such as wildfire search-and-rescue operations. To this end, it presents the first integration of large language model (LLM)-driven agent AI with edge computing, proposing three scalable and resilient edge-enabled deployment architectures that facilitate low-latency, highly autonomous multi-drone collaboration in mission-critical scenarios. Experimental results demonstrate that the proposed approach significantly improves search coverage, reduces mission completion time, and achieves higher levels of autonomy compared to conventional methods. These findings validate the effectiveness and practicality of synergistically combining LLM-based agents with edge computing for real-time disaster response applications.

0 citationsRead paper

A Dual-Purpose Framework for Backdoor Defense and Backdoor Amplification in Diffusion Models

Feb 26, 2025

Diffusion models are vulnerable to backdoor attacks—maliciously triggering harmful content generation upon injection of specific perturbations—yet existing defense and attack analysis methods remain fragmented and suboptimal. This paper proposes PureDiffusion, the first unified framework enabling *simultaneous* high-robustness backdoor detection and controllable attack enhancement. Its core innovation is a dual-loss trigger inversion reconstruction mechanism grounded in temporal distribution shift modeling and denoising consistency constraints, supporting bidirectional optimization. On the defense side, it achieves ≈100% detection accuracy—substantially surpassing state-of-the-art methods. On the attack side, lightweight trigger reinforcement training boosts attack success rates to ≈100% while reducing training time by 20×. PureDiffusion establishes a novel, interpretable, and reusable paradigm for backdoor security research in diffusion models.

0 citationsRead paper

Recent publications

Latest Papers

TooBad: Backdoor Diffusion Models with Ultra-Low Poison Rate and Imperceptible Trigger

Jun 22, 2026

Existing backdoor attacks on diffusion models struggle to simultaneously achieve high efficiency, low poisoning rates, and strong stealthiness. This work proposes TooBad, a novel framework that introduces, for the first time, a trigger optimization mechanism tailored specifically for diffusion models. By integrating fine-tuning with minimal poisoned data injection, TooBad attains over 85% attack success rate with only a 0.5% poisoning ratio; when the poisoning ratio increases to 5%, near-perfect success (≈100%) is achieved within just 3–5 training epochs. The method substantially reduces training overhead while effectively evading state-of-the-art defense mechanisms, thereby offering both potent attack performance and high concealment.

0 citationsRead paper

Critical-CoT: A Robust Defense Framework against Reasoning-Level Backdoor Attacks in Large Language Models

Apr 12, 2026

Large language models are vulnerable to reasoning-level backdoor attacks, wherein adversaries embed malicious reasoning steps into the chain-of-thought via trigger mechanisms, leading models to produce seemingly plausible yet harmful outputs. This work proposes Critical-CoT, a novel defense framework specifically designed to counter such attacks. Critical-CoT employs a two-stage fine-tuning strategy to endow models with critical thinking capabilities, enabling them to automatically detect and reject compromised reasoning steps. Experimental results demonstrate that the proposed method achieves strong robustness against both in-context learning and fine-tuning-based backdoor attacks across multiple mainstream large language models and datasets. Furthermore, it significantly enhances model safety while exhibiting excellent generalization across domains and tasks.

0 citationsRead paper

Agentic AI Meets Edge Computing in Autonomous UAV Swarms

Jan 20, 2026IEEE Internet of Things Magazine

This work addresses the challenge of achieving efficient and autonomous coordination among drone swarms in high-risk, infrastructure-constrained dynamic environments such as wildfire search-and-rescue operations. To this end, it presents the first integration of large language model (LLM)-driven agent AI with edge computing, proposing three scalable and resilient edge-enabled deployment architectures that facilitate low-latency, highly autonomous multi-drone collaboration in mission-critical scenarios. Experimental results demonstrate that the proposed approach significantly improves search coverage, reduces mission completion time, and achieves higher levels of autonomy compared to conventional methods. These findings validate the effectiveness and practicality of synergistically combining LLM-based agents with edge computing for real-time disaster response applications.

0 citationsRead paper

A Dual-Purpose Framework for Backdoor Defense and Backdoor Amplification in Diffusion Models

Feb 26, 2025

Diffusion models are vulnerable to backdoor attacks—maliciously triggering harmful content generation upon injection of specific perturbations—yet existing defense and attack analysis methods remain fragmented and suboptimal. This paper proposes PureDiffusion, the first unified framework enabling *simultaneous* high-robustness backdoor detection and controllable attack enhancement. Its core innovation is a dual-loss trigger inversion reconstruction mechanism grounded in temporal distribution shift modeling and denoising consistency constraints, supporting bidirectional optimization. On the defense side, it achieves ≈100% detection accuracy—substantially surpassing state-of-the-art methods. On the attack side, lightweight trigger reinforcement training boosts attack success rates to ≈100% while reducing training time by 20×. PureDiffusion establishes a novel, interpretable, and reusable paradigm for backdoor security research in diffusion models.

0 citationsRead paper