Wenqi Wei
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Wenqi Wei has a total of 97 co-authors. Here are the top co-authors by number of shared publications:
- Ling Liu
School of Computer Science, Georgia Institute of Technology
20 shared publications - Ka-Ho Chow
The University of Hong Kong
12 shared publications - Yanzhao Wu
Florida International University, Miami, FL, USA
9 shared publications - Jingya Zhou
School of Computer Science and Technology, Soochow University, Suzhou, China
5 shared publications - Mehmet Emre Gursoy
Koc University
5 shared publications - Stacey Truex
Georgia Institute of Technology 30332 Atlanta GA USA
5 shared publications
Latest Publications
- Open Access
- 22 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 145-158
Measuring the Robustness of Audio Deepfake Detection under Real-World Corruption
Deepfakes have emerged as a widespread and rapidly escalating concern in generative AI, spanning various media types such as images, audio, and videos. Among these, audio deepfakes are particularly alarming due to the growing accessibility of high-quality voice synthesis tools and the ease with which synthetic speech can be distributed via platforms like social media and robocalls. Consequently, detecting audio deepfakes is critical in combating the misuse of AI-synthesized speech. However, real-world audio is often subject to various corruptions, such as noise, modification, and compression, that may significantly impact detection performance. In this work, we systematically evaluate the robustness of 10 audio deepfake detection models against 18 common corruption types, grouped into categories: noise perturbation, audio modification, and compression. Using both traditional deep learning models and state-of-the-art foundation models, our study yields four key insights. (1) Most models demonstrate strong robustness to noise but they are notably more vulnerable to audio modifications and compression, especially when neural codecs are applied. (2) Speech foundation models generally outperform traditional models across most corruption scenarios, likely due to their extensive pre-training on large-scale and diverse audio datasets. (3) Increasing model size improves robustness, though with diminishing returns. (4) Robustness to unseen corruptions can be enhanced by targeted data augmentation during training or by applying speech enhancement techniques at inference time. These findings highlight the importance of comprehensive evaluation against diverse corruption types and developing more robust audio deepfake detection frameworks to ensure reliability in practical deployment settings. We further advocate that future research in deepfake detection across all media formats should account for the diverse and often unpredictable distortions common in real-world environments.
- 22 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 145-158
Open AccessMeasuring the Robustness of Audio Deepfake Detection under Real-World Corruption
Deepfakes have emerged as a widespread and rapidly escalating concern in generative AI, spanning various media types such as images, audio, and videos. Among these, audio deepfakes are particularly alarming due to the growing accessibility of high-quality voice synthesis tools and the ease with which synthetic speech can be distributed via platforms like social media and robocalls. Consequently, detecting audio deepfakes is critical in combating the misuse of AI-synthesized speech. However, real-world audio is often subject to various corruptions, such as noise, modification, and compression, that may significantly impact detection performance. In this work, we systematically evaluate the robustness of 10 audio deepfake detection models against 18 common corruption types, grouped into categories: noise perturbation, audio modification, and compression. Using both traditional deep learning models and state-of-the-art foundation models, our study yields four key insights. (1) Most models demonstrate strong robustness to noise but they are notably more vulnerable to audio modifications and compression, especially when neural codecs are applied. (2) Speech foundation models generally outperform traditional models across most corruption scenarios, likely due to their extensive pre-training on large-scale and diverse audio datasets. (3) Increasing model size improves robustness, though with diminishing returns. (4) Robustness to unseen corruptions can be enhanced by targeted data augmentation during training or by applying speech enhancement techniques at inference time. These findings highlight the importance of comprehensive evaluation against diverse corruption types and developing more robust audio deepfake detection frameworks to ensure reliability in practical deployment settings. We further advocate that future research in deepfake detection across all media formats should account for the diverse and often unpredictable distortions common in real-world environments.
- Open Access
- 11 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Internet Technology
- Vol. 26  (2) , 1-20
Improving Financial Statement Fraud Detection: A Large Language Model Processing Approach
With the prevalence of Internet AI technology, financial fraud becomes an imperative problem, especially in the context of machine learning. Technologies such as deep learning and natural language processing provide effective tools for detecting fraud with the guidance of financial statements, improving the efficiency and accuracy of data analysis, and helping to ensure financial safety. In this study, we propose a sophisticated representation learning method to detect financial statement fraud by tracking the detailed changes in the company’s Management Discussion and Analysis (MD&A) documents over time. Unlike traditional word frequency methods, we align paragraphs between consecutive disclosures based on their similarity at the paragraph level and categorize them into three types: added, deleted, and matched. Next, we create multivariate change trajectory representations based on fraud-related word categories. Finally, we use these word-level change trajectories to design a fraud detection model and compare it with several traditional models as well as the latest Time-Series Foundation Models. Experiments on 24 years of financial report data, from 1995 to 2019, show that our representation learning method significantly improves the performance of financial statement fraud detection across 11 different machine learning models, consistently outperforming traditional word frequency methods. Our method opens a new paradigm for feature engineering in financial statement fraud detection. Our code can be found at https://github.com/LittelStudent/Financial-Statement-Fraud-Detection-ParaEmb-FraudW2V .
- 11 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Internet Technology
- Vol. 26  (2) , 1-20
Open AccessImproving Financial Statement Fraud Detection: A Large Language Model Processing Approach
With the prevalence of Internet AI technology, financial fraud becomes an imperative problem, especially in the context of machine learning. Technologies such as deep learning and natural language processing provide effective tools for detecting fraud with the guidance of financial statements, improving the efficiency and accuracy of data analysis, and helping to ensure financial safety. In this study, we propose a sophisticated representation learning method to detect financial statement fraud by tracking the detailed changes in the company’s Management Discussion and Analysis (MD&A) documents over time. Unlike traditional word frequency methods, we align paragraphs between consecutive disclosures based on their similarity at the paragraph level and categorize them into three types: added, deleted, and matched. Next, we create multivariate change trajectory representations based on fraud-related word categories. Finally, we use these word-level change trajectories to design a fraud detection model and compare it with several traditional models as well as the latest Time-Series Foundation Models. Experiments on 24 years of financial report data, from 1995 to 2019, show that our representation learning method significantly improves the performance of financial statement fraud detection across 11 different machine learning models, consistently outperforming traditional word frequency methods. Our method opens a new paradigm for feature engineering in financial statement fraud detection. Our code can be found at https://github.com/LittelStudent/Financial-Statement-Fraud-Detection-ParaEmb-FraudW2V .
- 16 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Services Computing
- Vol. 19  (2) , 1277-1290
GradCloak: Gradient Obfuscation for Privacy-Preserving Distributed Learning as a Service
Gradient leakage attacks pose a significant privacy threat in distributed learning-as-a-service APIs. Existing literature on gradient leakage defense relies on gradient perturbation for preventing privacy leakage. However, determining where and how much to perturb the gradient offers different capabilities for preventing gradient leakage. This paper presents GradCloak, a principled approach to guiding gradient perturbation with theoretical robustness bounds in federated learning as a service, aiming to find the minimum required noise for simultaneously achieving privacy protection, competitive accuracy, and preventing gradient leakage attacks. The paper is organized into three major components. First, we formulate the gradient leakage threats and their adverse effect. We categorize the attack into two broad types: leakage during local training and leakage before global aggregation. Second, we investigate different gradient perturbation approaches. We analyze and compare these gradient perturbation methods, which are performed at the federated server, with those performed at the participating client(s). Third, we introduce three robustness properties of robust perturbation against gradient leakage threats, formulated by the anonymization bound for training data robustness, the perturbation bound for gradient robustness, and the distribution robustness bound for perturbed gradients. We conduct extensive evaluations on eight benchmark datasets to demonstrate that specific settings of gradient perturbation exist that best balance privacy, accuracy, and leakage prevention. Code is available at https://github.com/git-disl/GradCloak.
- 16 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Services Computing
- Vol. 19  (2) , 1277-1290
GradCloak: Gradient Obfuscation for Privacy-Preserving Distributed Learning as a Service
Gradient leakage attacks pose a significant privacy threat in distributed learning-as-a-service APIs. Existing literature on gradient leakage defense relies on gradient perturbation for preventing privacy leakage. However, determining where and how much to perturb the gradient offers different capabilities for preventing gradient leakage. This paper presents GradCloak, a principled approach to guiding gradient perturbation with theoretical robustness bounds in federated learning as a service, aiming to find the minimum required noise for simultaneously achieving privacy protection, competitive accuracy, and preventing gradient leakage attacks. The paper is organized into three major components. First, we formulate the gradient leakage threats and their adverse effect. We categorize the attack into two broad types: leakage during local training and leakage before global aggregation. Second, we investigate different gradient perturbation approaches. We analyze and compare these gradient perturbation methods, which are performed at the federated server, with those performed at the participating client(s). Third, we introduce three robustness properties of robust perturbation against gradient leakage threats, formulated by the anonymization bound for training data robustness, the perturbation bound for gradient robustness, and the distribution robustness bound for perturbed gradients. We conduct extensive evaluations on eight benchmark datasets to demonstrate that specific settings of gradient perturbation exist that best balance privacy, accuracy, and leakage prevention. Code is available at https://github.com/git-disl/GradCloak.
- Open Access
Quantum-Audit: Evaluating the Reasoning Limits of LLMs on Quantum Computing
Language models have become practical tools for quantum computing education and research, from summarizing technical papers to explaining theoretical concepts and answering questions about recent developments in the field. While existing benchmarks evaluate quantum code generation and circuit design, their understanding of quantum computing concepts has not been systematically measured. Quantum-Audit addresses this gap with 2,700 questions covering core quantum computing topics. We evaluate 26 models from leading organizations. Our benchmark comprises 1,000 expert-written questions, 1,000 questions extracted from research papers using LLMs and validated by experts, plus an additional 700 questions including 350 open-ended questions and 350 questions with false premises to test whether models can correct erroneous assumptions. Human participants scored between 23% and 86%, with experts averaging 74%. Top-performing models exceeded the expert average, with Claude Opus 4.5 reaching 84% accuracy, though top models showed an average 12-point accuracy drop on expert-written questions compared to LLM-generated ones. Performance declined further on advanced topics, dropping to 73% on security questions. Additionally, models frequently accepted and reinforced false premises embedded in questions instead of identifying them, with accuracy below 66% on these critical reasoning tasks.
Open AccessQuantum-Audit: Evaluating the Reasoning Limits of LLMs on Quantum Computing
Language models have become practical tools for quantum computing education and research, from summarizing technical papers to explaining theoretical concepts and answering questions about recent developments in the field. While existing benchmarks evaluate quantum code generation and circuit design, their understanding of quantum computing concepts has not been systematically measured. Quantum-Audit addresses this gap with 2,700 questions covering core quantum computing topics. We evaluate 26 models from leading organizations. Our benchmark comprises 1,000 expert-written questions, 1,000 questions extracted from research papers using LLMs and validated by experts, plus an additional 700 questions including 350 open-ended questions and 350 questions with false premises to test whether models can correct erroneous assumptions. Human participants scored between 23% and 86%, with experts averaging 74%. Top-performing models exceeded the expert average, with Claude Opus 4.5 reaching 84% accuracy, though top models showed an average 12-point accuracy drop on expert-written questions compared to LLM-generated ones. Performance declined further on advanced topics, dropping to 73% on security questions. Additionally, models frequently accepted and reinforced false premises embedded in questions instead of identifying them, with accuracy below 66% on these critical reasoning tasks.
- Open Access
- 15 January 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Internet Technology
- Vol. 26  (1) , 1-2
Introduction to the Special Issue on Reliable Infrastructure and Edge Analytics for IoT
- 15 January 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Internet Technology
- Vol. 26  (1) , 1-2
Open AccessIntroduction to the Special Issue on Reliable Infrastructure and Edge Analytics for IoT
- Open Access
FastFHE: Packing-Scalable and Depthwise-Separable CNN Inference Over FHE
The deep learning (DL) has been penetrating daily life in many domains, how to keep the DL model inference secure and sample privacy in an encrypted environment has become an urgent and increasingly important issue for various security-critical applications. To date, several approaches have been proposed based on the Residue Number System variant of the Cheon-Kim-Kim-Song (RNS-CKKS) scheme. However, they all suffer from high latency, which severely limits the applications in real-world tasks. Currently, the research on encrypted inference in deep CNNs confronts three main bottlenecks: i) the time and storage costs of convolution calculation; ii) the time overhead of huge bootstrapping operations; and iii) the consumption of circuit multiplication depth. Towards these three challenges, we in this paper propose an efficient and effective mechanism FastFHE to accelerate the model inference while simultaneously retaining high inference accuracy over fully homomorphic encryption. Concretely, our work elaborates four unique novelties. First, we propose a new scalable ciphertext data-packing scheme to save the time and storage consumptions. Second, we work out a depthwise-separable convolution fashion to degrade the computation load of convolution calculation. Third, we figure out a BN dot-product fusion matrix to merge the ciphertext convolutional layer with the batch-normalization layer without incurring extra multiplicative depth. Last but not least, we adopt the low-degree Legendre polynomial to approximate the nonlinear smooth activation function SiLU under the guarantee of tiny accuracy error before and after encrypted inference. Finally, we execute multi-facet experiments to verify the efficiency and effectiveness of our proposed approach.
Open AccessFastFHE: Packing-Scalable and Depthwise-Separable CNN Inference Over FHE
The deep learning (DL) has been penetrating daily life in many domains, how to keep the DL model inference secure and sample privacy in an encrypted environment has become an urgent and increasingly important issue for various security-critical applications. To date, several approaches have been proposed based on the Residue Number System variant of the Cheon-Kim-Kim-Song (RNS-CKKS) scheme. However, they all suffer from high latency, which severely limits the applications in real-world tasks. Currently, the research on encrypted inference in deep CNNs confronts three main bottlenecks: i) the time and storage costs of convolution calculation; ii) the time overhead of huge bootstrapping operations; and iii) the consumption of circuit multiplication depth. Towards these three challenges, we in this paper propose an efficient and effective mechanism FastFHE to accelerate the model inference while simultaneously retaining high inference accuracy over fully homomorphic encryption. Concretely, our work elaborates four unique novelties. First, we propose a new scalable ciphertext data-packing scheme to save the time and storage consumptions. Second, we work out a depthwise-separable convolution fashion to degrade the computation load of convolution calculation. Third, we figure out a BN dot-product fusion matrix to merge the ciphertext convolutional layer with the batch-normalization layer without incurring extra multiplicative depth. Last but not least, we adopt the low-degree Legendre polynomial to approximate the nonlinear smooth activation function SiLU under the guarantee of tiny accuracy error before and after encrypted inference. Finally, we execute multi-facet experiments to verify the efficiency and effectiveness of our proposed approach.
- Open Access
An Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs
With the remarkable success of Vision-Language Models (VLMs) on multimodal tasks, concerns regarding their deployment efficiency have become increasingly prominent. In particular, the number of tokens consumed during the generation process has emerged as a key evaluation metric.Prior studies have shown that specific inputs can induce VLMs to generate lengthy outputs with low information density, which significantly increases energy consumption, latency, and token costs. However, existing methods simply delay the occurrence of the EOS token to implicitly prolong output, and fail to directly maximize the output token length as an explicit optimization objective, lacking stability and controllability.To address these limitations, this paper proposes a novel verbose-text induction attack (VTIA) to inject imperceptible adversarial perturbations into benign images via a two-stage framework, which identifies the most malicious prompt embeddings for optimizing and maximizing the output token of the perturbed images.Specifically, we first perform adversarial prompt search, employing reinforcement learning strategies to automatically identify adversarial prompts capable of inducing the LLM component within VLMs to produce verbose outputs. We then conduct vision-aligned perturbation optimization to craft adversarial examples on input images, maximizing the similarity between the perturbed image's visual embeddings and those of the adversarial prompt, thereby constructing malicious images that trigger verbose text generation. Comprehensive experiments on four popular VLMs demonstrate that our method achieves significant advantages in terms of effectiveness, efficiency, and generalization capability.
Open AccessAn Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs
With the remarkable success of Vision-Language Models (VLMs) on multimodal tasks, concerns regarding their deployment efficiency have become increasingly prominent. In particular, the number of tokens consumed during the generation process has emerged as a key evaluation metric.Prior studies have shown that specific inputs can induce VLMs to generate lengthy outputs with low information density, which significantly increases energy consumption, latency, and token costs. However, existing methods simply delay the occurrence of the EOS token to implicitly prolong output, and fail to directly maximize the output token length as an explicit optimization objective, lacking stability and controllability.To address these limitations, this paper proposes a novel verbose-text induction attack (VTIA) to inject imperceptible adversarial perturbations into benign images via a two-stage framework, which identifies the most malicious prompt embeddings for optimizing and maximizing the output token of the perturbed images.Specifically, we first perform adversarial prompt search, employing reinforcement learning strategies to automatically identify adversarial prompts capable of inducing the LLM component within VLMs to produce verbose outputs. We then conduct vision-aligned perturbation optimization to craft adversarial examples on input images, maximizing the similarity between the perturbed image's visual embeddings and those of the adversarial prompt, thereby constructing malicious images that trigger verbose text generation. Comprehensive experiments on four popular VLMs demonstrate that our method achieves significant advantages in terms of effectiveness, efficiency, and generalization capability.
- Open Access
- 14 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Quantum Engineering
- Vol. 7, 1-15
Hardware-Aware and Resource-Efficient Circuit Packing and Scheduling on Trapped-Ion Quantum Computers
The rapid expansion of quantum cloud services has led to long job queues due to single-tenant execution models that underutilize hardware resources. Quantum multiprogramming (QMP) mitigates this by executing multiple circuits in parallel on a single device, but existing methods target superconducting systems with limited connectivity, high crosstalk, and lower gate fidelity. Trapped-ion architecture, with all-to-all connectivity, long coherence times, and high-fidelity mid-circuit measurement properties, presents itself as a more suitable platform for scalable QMP. We present CircPack, a hardware-aware circuit packing framework designed for modular trapped-ion devices based on the quantum charge-coupled device (QCCD) architecture. CircPack formulates static circuit scheduling as a 2-D packing problem with hardware-specific shuttling constraints. Compared to superconducting-based QMP approaches, CircPack achieves up to 70.72% better fidelity, 62.67% higher utilization, and 32.80% improved layer reduction. This framework is also capable of scalable balanced scheduling across a cluster of independent QCCD modules, highlighting trapped-ion systems’ potential in improving the throughput of quantum cloud computing in the near future.
- 14 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Quantum Engineering
- Vol. 7, 1-15
Open AccessHardware-Aware and Resource-Efficient Circuit Packing and Scheduling on Trapped-Ion Quantum Computers
The rapid expansion of quantum cloud services has led to long job queues due to single-tenant execution models that underutilize hardware resources. Quantum multiprogramming (QMP) mitigates this by executing multiple circuits in parallel on a single device, but existing methods target superconducting systems with limited connectivity, high crosstalk, and lower gate fidelity. Trapped-ion architecture, with all-to-all connectivity, long coherence times, and high-fidelity mid-circuit measurement properties, presents itself as a more suitable platform for scalable QMP. We present CircPack, a hardware-aware circuit packing framework designed for modular trapped-ion devices based on the quantum charge-coupled device (QCCD) architecture. CircPack formulates static circuit scheduling as a 2-D packing problem with hardware-specific shuttling constraints. Compared to superconducting-based QMP approaches, CircPack achieves up to 70.72% better fidelity, 62.67% higher utilization, and 32.80% improved layer reduction. This framework is also capable of scalable balanced scheduling across a cluster of independent QCCD modules, highlighting trapped-ion systems’ potential in improving the throughput of quantum cloud computing in the near future.
- 13 October 2025
- Published by Springer Nature
- p. 388-409
On the Adversarial Robustness of Graph Neural Networks with Graph Reduction
- 13 October 2025
- Published by Springer Nature
- p. 388-409
On the Adversarial Robustness of Graph Neural Networks with Graph Reduction
- Open Access
- 18 August 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Internet Technology
- Vol. 25  (3) , 1-19
Where are We in Audio Deepfake Detection? A Systematic Analysis over Generative and Detection Models
Recent advances in Text-to-Speech (TTS) and Voice-Conversion (VC) using generative Artificial Intelligence (AI) technology have made it possible to generate high-quality and realistic human-like audio. This introduces significant challenges to distinguishing AI-synthesized speech from the authentic human voice and could raise potential issues of misuse for malicious purposes such as impersonation and fraud, spreading misinformation, deepfakes, and scams. However, existing detection techniques for AI-synthesized audio have not kept pace and often exhibit poor generalization across diverse datasets. In this article, we introduce SONAR, a s ynthetic AI-Audi o Detectio n Fr a mework and Benchma r k, aiming at providing a comprehensive evaluation for distinguishing cutting-edge AI-synthesized auditory content. SONAR includes a novel evaluation dataset sourced from 9 diverse audio synthesis platforms, including leading TTS providers and state-of-the-art TTS models. It is the first framework to uniformly benchmark AI-audio detection across both traditional and foundation model-based deepfake detection systems. Through extensive experiments, (1) we reveal the generalization limitations of existing detection methods and demonstrate that foundation models exhibit stronger generalization capabilities, which can be attributed to their model size and the scale and quality of pretraining data. (2) Our evaluation of the generalization across languages suggests that speech foundation models demonstrate robust cross-lingual generalization capabilities, maintaining strong performance across diverse languages despite being fine-tuned solely on English speech data. This finding also suggests that the primary challenges in audio deepfake detection are more closely tied to the realism and quality of synthetic audio rather than language-specific characteristics. (3) We also explore the effectiveness and efficiency of few-shot fine-tuning in improving generalization, highlighting its potential for tailored applications, such as personalized detection systems for specific entities or individuals. Code and dataset are available at https://github.com/Jessegator/SONAR .
- 18 August 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Internet Technology
- Vol. 25  (3) , 1-19
Open AccessWhere are We in Audio Deepfake Detection? A Systematic Analysis over Generative and Detection Models
Recent advances in Text-to-Speech (TTS) and Voice-Conversion (VC) using generative Artificial Intelligence (AI) technology have made it possible to generate high-quality and realistic human-like audio. This introduces significant challenges to distinguishing AI-synthesized speech from the authentic human voice and could raise potential issues of misuse for malicious purposes such as impersonation and fraud, spreading misinformation, deepfakes, and scams. However, existing detection techniques for AI-synthesized audio have not kept pace and often exhibit poor generalization across diverse datasets. In this article, we introduce SONAR, a s ynthetic AI-Audi o Detectio n Fr a mework and Benchma r k, aiming at providing a comprehensive evaluation for distinguishing cutting-edge AI-synthesized auditory content. SONAR includes a novel evaluation dataset sourced from 9 diverse audio synthesis platforms, including leading TTS providers and state-of-the-art TTS models. It is the first framework to uniformly benchmark AI-audio detection across both traditional and foundation model-based deepfake detection systems. Through extensive experiments, (1) we reveal the generalization limitations of existing detection methods and demonstrate that foundation models exhibit stronger generalization capabilities, which can be attributed to their model size and the scale and quality of pretraining data. (2) Our evaluation of the generalization across languages suggests that speech foundation models demonstrate robust cross-lingual generalization capabilities, maintaining strong performance across diverse languages despite being fine-tuned solely on English speech data. This finding also suggests that the primary challenges in audio deepfake detection are more closely tied to the realism and quality of synthetic audio rather than language-specific characteristics. (3) We also explore the effectiveness and efficiency of few-shot fine-tuning in improving generalization, highlighting its potential for tailored applications, such as personalized detection systems for specific entities or individuals. Code and dataset are available at https://github.com/Jessegator/SONAR .