Publications

Papers

Journal papers 7

Conference papers 18

Preprints 4

Citations

Citations 311

h-index 7

i10-index 6

Lines show papers and citation counts by publication year. A year with no papers or citations is drawn as zero. The list is refreshed from OpenAlex and Semantic Scholar for ORCID 0000-0001-6583-9526 (last checked 2026-11-13). The citation total, h-index, and i10-index are the OpenAlex author counts.

# Title Venue
1
Figure 3

Figure 3

Decide the next psychiatric question from a criteria-grounded evidence state.

Each turn stores observed evidence, open criterion checks, active differentials, and how reliable the support is. That shared state selects the action and scores information gain.

EMNLP 2026 2026 1
2 The Local–Global Alignment Paradox: Evaluating Distributed Safety Reconstruction in Multi-Agent Systems UKAIRS 26 2026 0
3
Figure 2

Figure 2

Caption open-world images using a knowledge graph aligned to the scene.

AKGMA links visual entities to a knowledge graph, builds an image-specific subgraph, and adds scene context before generation. The aim is fewer knowledge hallucinations and more specific captions.

AAAI 2026 2026 0
4
Figure 2

Figure 2

Measure social bias in LLM-based video game NPCs.

The benchmark scores class, race, age, and nationality bias in transaction, cooperation, and competition. Across 12 tasks, larger models showed stronger bias.

ACL 2026 2026 0
5 PrivacyShadow: Revealing Fine-Tuning Leakage in Vision-Language Models Via Dual-Level Black-Box Attacks ICASSP 2026 2026 0
6
Figure 1

Figure 1

Measure how much structure a black-box PEFT adapter leaks.

VectorHijack-SR compares local runs of the public base model with the service output. From those residuals it bounds the adapter family, layer, and coarse rank.

arXiv 2026 1
7
Figure 3

Figure 3

Decide the next psychiatric question from a criteria-grounded evidence state.

Each turn stores observed evidence, open criterion checks, active differentials, and how reliable the support is. That shared state selects the action and scores information gain.

arXiv 2026 0
8
Figure 2

Figure 2

Measure social bias in LLM-based video game NPCs.

The benchmark scores class, race, age, and nationality bias in transaction, cooperation, and competition. Across 12 tasks, larger models showed stronger bias.

arXiv 2025 0
9 Enhancing LLM Jailbreak: A Dual-Strategy Approach of Token Suppression and Induction CSCWD 2025 2025 2
10 Emotion-aware Structural Enhancement Graph Auto-Encoder for Rumor Detection ICASSP 2025 2025 3
11
Figure 1

Figure 1

Jailbreak an LLM by hiding the query's real intent.

The attack either obscures the intent or makes the query ambiguous so the model's checks miss it. On ChatGPT, Qwen, and Baichuan it succeeded for 69.21% of prompts.

The Computer Journal 2024 19
12
Figure 1

Figure 1

Jailbreak an LLM by hiding the query's real intent.

The attack either obscures the intent or makes the query ambiguous so the model's checks miss it. On ChatGPT, Qwen, and Baichuan it succeeded for 69.21% of prompts.

arXiv 2024 1
13
Figure 3

Figure 3

Tune jailbreak prompts with a reinforcement-learning reward.

The reward trades off readability, diversity, and attack strength. On six open models and three commercial APIs, the prompts beat mainstream jailbreaks on attack success rate.

CSAI 2024 2024 4
14 IntentObfuscator: A jailbreaking method via confuses LLM with prompts ESORICS 2024 2024 6
15 EthGAN: Improving Ethereum Account Classification Accurary via Data Augmentation IJCNN 2024 2024 4
16 Reinforcement Learning-Powered Effectiveness and Effciency Few-Shot Jailbreaking Attack LLMs IEEE ISPA 2024 2024 2
17
Figure 3

Figure 3

Classify blockchain consensus by the agreement process.

The survey traces how representative algorithms evolved and groups them by that process. It then compares them on decentralization, scalability, security, consistency, availability, and partition tolerance.

Journal of Cyber Security 2021 —
18
Figure 1

Figure 1

Classify encrypted flows from packet size and inter-packet time.

A Gaussian mixture clusters those two features, and a hidden Markov model treats the clusters as observations of the hidden protocol state.

JNCA 2020 62
19
Figure 2

Figure 2

Vouch for a packet without learning who sent or received it.

Sender and receiver use a temporary anonymous identity with a communication agent. The agent checks that identity and still does not learn the real owner.

JPDC 2019 14
20 An Invisible Flow Watermarking for Traffic Tracking: A Hidden Markov Model Approach ICC 2019 8
21 A New Bitcoin Address Association Method Using a Two-Level Learner Model ICA3PP 2019 14
22 Encrypted Traffic Classification with a Convolutional Long Short-Term Memory Neural Network HPCC/SmartCity/DSS 2018 151
23 Meek-Based Tor Traffic Identification with Hidden Markov Model HPCC/SmartCity/DSS 2018 19
24 Research Review on Traffic Obfuscation and Its Corresponding Identification and Tracking Technologies Journal of Software 2018 —
25 A Summary of the Theory and Application of BlockChain e-Science Technology & Application 2017 —
26 Performance Evaluation and Analysis of the Virtual Networks Built by Different Virtual Machine Managers HPCC/SmartCity/DSS 2016 0
27 Performance Investigation and Comparison between Virtual Networks and Physical Networks Based on an Advanced Testbed Network CIT/IUCC/DASC/PICom 2015 0
28 DESIGN AND IMPLEMENTATION OF IS-IS NETWORK-ORIENTED NETMONITOR SYSTEM Computer Applications and Software 2015 —
29
Fig. 1

Fig. 1

Stop invalid routes from being injected into OSPF and BGP.

Injecting invalid routes from stub software routers raises the resource use of other routers. The paper then gives checks that reject those routes.

IIP 2014 1