Author name cluster

Li Pan

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

7 papers

1 author row

EAAI Journal 2026 Journal Article

A risk assessment framework for online transactions via Graph Neural Networks and efficient probabilistic prediction

Jicai Chang
Xuejing Fu
Zhen Chen
Li Pan
Shijun Liu

Transactions are integral to daily life, but the occurrence of abnormal behaviors can lead to significant risks. Online transaction risk is characterized by the accumulation of abnormal behaviors, where their frequency surpasses a predefined threshold, resulting in measurable probabilities and consequences. Therefore, the assessment of online transaction risk heavily depends on probabilistic predictions of the accumulated frequency of abnormal behaviors, presenting two major challenges. Firstly, abnormal behaviors across different instances (e. g. , behavior types, product categories, regions, and platforms) exhibit temporal correlations, such as co-occurrence and concomitance, which most probabilistic models fail to identify and utilize effectively. Additionally, these models do not fully address the real-time demands. To address these challenges, we propose a novel risk assessment framework based on Graph Neural Network (GNN) and probabilistic prediction, named GNN-Probformer. The framework uses Dynamic Time Warping to capture temporal correlations between abnormal behavior frequency sequences and constructs a graph structure through clustering. It then employs Graph Neural Networks to aggregate features and learn representations through a novel embedding module. A sparse self-attention mechanism and an efficient encoder–decoder architecture are incorporated to further enhance performance, while probabilistic predictions are generated through Monte Carlo sampling and cumulative distribution functions. Experimental results on a real-world dataset demonstrate that GNN-Probformer achieves substantial performance gains, with a 15% reduction in normalized deviation. At the 90th percentile, it further reduces normalized quantile loss by 15% and improves the F1-score by 16%, while also reducing training time and inference time by 47% and 38%, respectively.

Details DOI

AAAI Conference 2026 Conference Paper

SceneJailEval: A Scenario-Adaptive Multi-Dimensional Framework for Jailbreak Evaluation

Lai Jiang
Yuekang Li
Xiaohan Zhang
Youtao Ding
Li Pan

Accurate jailbreak evaluation is critical for LLM red team testing and jailbreak research. Mainstream methods rely on binary classification (string matching, toxic text classifiers, and LLM-based methods), outputting only "yes/no" labels without quantifying harm severity. Emerged multi-dimensional frameworks (e.g., Security Violation, Relative Truthfulness and Informativeness) use unified evaluation standards across scenarios, leading to scenario-specific mismatches (e.g., "Relative Truthfulness" is irrelevant to "hate speech"), undermining evaluation accuracy. To address these, we propose SceneJailEval, with key contributions: (1) A pioneering scenario-adaptive multi-dimensional framework for jailbreak evaluation, overcoming the critical "one-size-fits-all" limitation of existing multi-dimensional methods, and boasting robust extensibility to seamlessly adapt to customized or emerging scenarios. (2) A novel 14-scenario dataset featuring rich jailbreak variants and regional cases, addressing the long-standing gap in high-quality, comprehensive benchmarks for scenario-adaptive evaluation. (3) SceneJailEval delivers state-of-the-art performance with an F1 score of 0.917 on our full-scenario dataset (+6% over SOTA) and 0.995 on JBB (+3% over SOTA), breaking through the accuracy bottleneck of existing evaluation methods in heterogeneous scenarios and solidifying its superiority.

PDF Details DOI

AAAI Conference 2026 Conference Paper

Your Prompts Are Not Safe: Output-Free Membership Inference via Prompt Vectors in Vision-Language Tuning

Yuran Bian
Xiaohan Zhang
Zhiyuan Yu
Changqing Li
Li Pan

Prompt tuning enables Vision-Language Models (VLMs) to efficiently adapt to new tasks through learnable prompt vectors. This naturally raises a question: do these prompts leak private information about their training data? While Membership Inference Attacks (MIAs) can quantify this risk, current methods rely on access to model outputs or internal gradients. This limitation prevents a clear assessment of a prompt’s standalone privacy leakage, particularly in deployment scenarios where such information is inaccessible. In this paper, we propose Prompt Intrinsic Privacy Risk Analyzer (PIPRA) to address this gap. As the first output-free MIA, PIPRA leverages open-source pre-trained VLMs to extract features from both prompts and samples within a shared cross-modal semantic space. By employing a contrastive learning-based feature projector to enhance these representations, PIPRA enables a subsequent discriminator to effectively perform membership inference. Extensive experiments across nine benchmark datasets and multiple VLMs show PIPRA achieves an average AUC of 87.58%, significantly outperforming traditional output-dependent methods (77.05%). These findings reveal that prompts pose a substantially greater privacy risk than previously recognized, highlighting the urgent need for prompt-level privacy protection.

PDF Details DOI

JBHI Journal 2025 Journal Article

Unified Multi-Modal Diagnostic Framework With Reconstruction Pre-Training and Heterogeneity-Combat Tuning

Yupei Zhang
Li Pan
Qiushi Yang
Tan Li
Zhen Chen

Medical multi-modal pre-training has revealed promise in computer-aided diagnosis by leveraging large-scale unlabeled datasets. However, existing methods based on masked autoencoders mainly rely on data-level reconstruction tasks, but lack high-level semantic information. Furthermore, two significant heterogeneity challenges hinder the transfer of pre-trained knowledge to downstream tasks, i. e. , the distribution heterogeneity between pre-training data and downstream data, and the modality heterogeneity within downstream data. To address these challenges, we propose a Unified Medical Multi-modal Diagnostic (UMD) framework with tailored pre-training and downstream tuning strategies. Specifically, to enhance the representation abilities of vision and language encoders, we propose the Multi-level Reconstruction Pre-training (MR-Pretrain) strategy, including a feature-level and data-level reconstruction, which guides models to capture the semantic information from masked inputs of different modalities. Moreover, to tackle two kinds of heterogeneities during the downstream tuning, we present the heterogeneity-combat downstream tuning strategy, which consists of a Task-oriented Distribution Calibration (TD-Calib) and a Gradient-guided Modality Coordination (GM-Coord). In particular, TD-Calib fine-tunes the pre-trained model regarding the distribution of downstream datasets, and GM-Coord adjusts the gradient weights according to the dynamic optimization status of different modalities. Extensive experiments on five public medical datasets demonstrate the effectiveness of our UMD framework, which remarkably outperforms existing approaches on three kinds of downstream tasks.

Details DOI

EAAI Journal 2023 Journal Article

An optimal allocation method for power distribution network partitions based on improved spectral clustering algorithm

Li Pan
Zhang Han
Zhao Shanshan
Wang Feng

Distribution network nodes are numerous and monitoring devices are widely distributed. All monitoring data are uploaded to the cloud master for centralized processing may cause serious problems, such as network congestion, information delay and high computational complexity. The edge computing can provide a good solution, which requires reasonable distribution network partition. This paper proposes a Monte Carlo optimized spectral clustering (MCOSC) distribution network partition method for edge server configuration. Firstly, the objective function of distribution network partition number is constructed with economic and real-time communication indexes, which are more suitable for edge computing than electrical distance and voltage sensitivity indexes. Then the optimal number of partitions is obtained by particle swarm optimization (PSO) with the constraints of communication reliability. Secondly, to solve the problem that the traditional spectral clustering is easy to fall into the local optimal solution, a Monte Carlo optimized spectral clustering method is proposed to make the distribution network partition results more reasonable. Finally, the performance of the proposed method is evaluated by IEEE 33 bus and 69 bus systems distribution network models. The results indicate that the Monte Carlo optimization partition method has better accurate, robustness and convergence speed than traditional spectral clustering method.

Details DOI

AAMAS Conference 2019 Conference Paper

Hybrid BiLSTM-Siamese Network for Relation Extraction

Zeyuan Cui
Li Pan
Shijun Liu

Relation extraction is an important processing task in knowledge graph completion. In previous approaches, it is considered to be a multi-class classification problem. In this paper, we propose a novel approach called hybrid BiLSTM-Siamese network which combines two word-level bidirectional LSTMs by a Siamese model architecture. It learns a similarity metric between two sentences and predicts the relation of a new sentence by k-nearest neighbors algorithm. In experiments, we use the SemEval-2010 Task8 dataset and achieve an F1-score of 81. 8%.

PDF

YNIMG Journal 2019 Journal Article

Multi-vendor standardized sequence for edited magnetic resonance spectroscopy

Muhammad G. Saleh
Daniel Rimbault
Mark Mikkelsen
Georg Oeltzschner
Anna M. Wang
Dengrong Jiang
Ali Alhamud
Jamie Near

Spectral editing allows direct measurement of low-concentration metabolites, such as GABA, glutathione (GSH) and lactate (Lac), relevant for understanding brain (patho)physiology. The most widely used spectral editing technique is MEGA-PRESS, which has been diversely implemented across research sites and vendors, resulting in variations in the final resolved edited signal. In this paper, we describe an effort to develop a new universal MEGA-PRESS sequence with HERMES functionality for the major MR vendor platforms with standardized RF pulse shapes, durations, amplitudes and timings. New RF pulses were generated for the universal sequence. Phantom experiments were conducted on Philips, Siemens, GE and Canon 3 T MRI scanners using 32-channel head coils. In vivo experiments were performed on the same six subjects on Philips and Siemens scanners, and on two additional subjects, one on GE and one on Canon scanners. On each platform, edited MRS experiments were conducted with the vendor-native and universal MEGA-PRESS sequences for GABA (TE = 68 ms) and Lac editing (TE = 140 ms). Additionally, HERMES for GABA and GSH was performed using the universal sequence at TE = 80 ms. The universal sequence improves inter-vendor similarity of GABA-edited and Lac-edited MEGA-PRESS spectra. The universal HERMES sequence yields both GABA- and GSH-edited spectra with negligible levels of crosstalk on all four platforms, and with strong agreement among vendors for both edited spectra. In vivo GABA+/Cr, Lac/Cr and GSH/Cr ratios showed relatively low variation between scanners using the universal sequence. In conclusion, phantom and in vivo experiments demonstrate successful implementation of the universal sequence across all four major vendors, allowing editing of several metabolites across a range of TEs.

Details DOI