Haiyang Yu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Haiyang Yu has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Ping Yi
Shanghai Jiao Tong University, Shanghai, China
3 shared publications - Yue Wu
Shanghai Jiao Tong University
3 shared publications - Futai Zou
School of Computer Science, Shanghai Jiao Tong University, Shanghai, China
2 shared publications - Jiaping Gui
Shanghai Jiao Tong University, Shanghai, China
2 shared publications - Fan Hong
School of Computer Science, Shanghai Jiao Tong University, Shanghai, China
1 shared publication - Futai Zou
School of Computer Science, Shanghai Jiao Tong University, China
1 shared publication
Latest Publications
- Open Access
- 15 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
FiCoVuL: A Framework for Fine-grained and Cross-function Code Vulnerability Detection
Detecting vulnerabilities in software development is crucial yet challenging. Deep learning-based approaches have shown promise in automatically learning features for vulnerable function detection. In practice, human analysts need to correlate the behavioral logic of multiple functions to confirm the occurrence of vulnerabilities. However, existing works fail to consider the contextual information across functions. Furthermore, human analysts expect models to provide finer-grained explanations to assist in the fixing process. In this paper, we present FiCoVuL, a framework for analyzing interconnected functions and providing fine-grained guidance for vulnerability fixing. FiCoVuL utilizes Function Fusion to extract function definitions and their call relationships from a project, and synthesize multiple functions into one for joint analysis. It models code as a multi-relational graph, capturing rich syntactic and semantic relationships between code statements. A multi-relational graph attention network is leveraged to perform message passing between nodes and edge relationships, generating graph-level and node-level representations. These representations are utilized for vulnerability function prediction and code line ranking, assisting in vulnerability fixing. We evaluate FiCoVuL on a public vulnerability dataset and a self-constructed cross-function one, and compare it with five existing detection models. Experimental results show that FiCoVuL significantly outperforms other methods in both vulnerability detection and localization.
- 15 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
Open AccessFiCoVuL: A Framework for Fine-grained and Cross-function Code Vulnerability Detection
Detecting vulnerabilities in software development is crucial yet challenging. Deep learning-based approaches have shown promise in automatically learning features for vulnerable function detection. In practice, human analysts need to correlate the behavioral logic of multiple functions to confirm the occurrence of vulnerabilities. However, existing works fail to consider the contextual information across functions. Furthermore, human analysts expect models to provide finer-grained explanations to assist in the fixing process. In this paper, we present FiCoVuL, a framework for analyzing interconnected functions and providing fine-grained guidance for vulnerability fixing. FiCoVuL utilizes Function Fusion to extract function definitions and their call relationships from a project, and synthesize multiple functions into one for joint analysis. It models code as a multi-relational graph, capturing rich syntactic and semantic relationships between code statements. A multi-relational graph attention network is leveraged to perform message passing between nodes and edge relationships, generating graph-level and node-level representations. These representations are utilized for vulnerability function prediction and code line ranking, assisting in vulnerability fixing. We evaluate FiCoVuL on a public vulnerability dataset and a self-constructed cross-function one, and compare it with five existing detection models. Experimental results show that FiCoVuL significantly outperforms other methods in both vulnerability detection and localization.
- Open Access
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Training-Time Detection of Fixed-Target Backdoor Data Poisons in Stable Diffusion Fine-Tuning via Anchored Denoising Loss Trajectories
Fixed-target data poisoning, in which an adversary injects image-caption pairs that all map to a single attacker-chosen target output, is a practical and attractive backdoor threat for Stable Diffusion fine-tuning, since its consistent and easily verifiable output appeals to attackers. However, existing defenses for fine-tuned diffusion models operate only at inference time on a deployed model, leaving the fine-tuning dataset itself unexamined and assuming the trigger pattern remains unknown. We identify a coordinated signature in the training dynamics of fine-tuning that is characteristic of this threat. As fixed-target memorization progresses, poisoned samples develop a markedly steeper denoising-loss decline than clean samples. Because they share the same target, these samples often cluster tightly together in trajectory space across many domains. To our knowledge, no prior defense has characterized or exploited this signature. Motivated by this finding, we propose ADLT, the first method that detects and removes fixed-target poisoned samples from a Stable Diffusion fine-tuning dataset by tracking this signature during training, without any knowledge of the trigger pattern or target image. ADLT records each sample's timestep-conditioned denoising loss at several fixed anchor timesteps across training epochs to form a per-sample trajectory, which is then scored by its Mahalanobis distance from a small set of trusted clean probe trajectories, with three complementary scoring modes adapting to different training dynamics. Experiments on Stable Diffusion and DDPM benchmarks show near-perfect detection, with best AUROC values of 1.000 on SD Pokemon, 0.983 on Oxford Flowers, and 0.9996 on DDPM. This work not only enables practical, trigger-agnostic protection of diffusion model fine-tuning from fixed-target backdoors but also points toward training dynamics as a broader lens for understanding dataset integrity in generative model training.
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessTraining-Time Detection of Fixed-Target Backdoor Data Poisons in Stable Diffusion Fine-Tuning via Anchored Denoising Loss Trajectories
Fixed-target data poisoning, in which an adversary injects image-caption pairs that all map to a single attacker-chosen target output, is a practical and attractive backdoor threat for Stable Diffusion fine-tuning, since its consistent and easily verifiable output appeals to attackers. However, existing defenses for fine-tuned diffusion models operate only at inference time on a deployed model, leaving the fine-tuning dataset itself unexamined and assuming the trigger pattern remains unknown. We identify a coordinated signature in the training dynamics of fine-tuning that is characteristic of this threat. As fixed-target memorization progresses, poisoned samples develop a markedly steeper denoising-loss decline than clean samples. Because they share the same target, these samples often cluster tightly together in trajectory space across many domains. To our knowledge, no prior defense has characterized or exploited this signature. Motivated by this finding, we propose ADLT, the first method that detects and removes fixed-target poisoned samples from a Stable Diffusion fine-tuning dataset by tracking this signature during training, without any knowledge of the trigger pattern or target image. ADLT records each sample's timestep-conditioned denoising loss at several fixed anchor timesteps across training epochs to form a per-sample trajectory, which is then scored by its Mahalanobis distance from a small set of trusted clean probe trajectories, with three complementary scoring modes adapting to different training dynamics. Experiments on Stable Diffusion and DDPM benchmarks show near-perfect detection, with best AUROC values of 1.000 on SD Pokemon, 0.983 on Oxford Flowers, and 0.9996 on DDPM. This work not only enables practical, trigger-agnostic protection of diffusion model fine-tuning from fixed-target backdoors but also points toward training dynamics as a broader lens for understanding dataset integrity in generative model training.
- 30 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP (15) , 1
SFBD: Backdoor Detection via Sequential Fingerprinting of Neural Networks for Securing the IoT Model Supply Chain
The rapid deployment of Deep Learning (DL) in the Internet of Things (IoT) increasingly relies on outsourcing model training to third parties or cloud servers due to the resource constraints of edge devices. In practical IoT scenarios, edge devices and gateways typically do not have access to original training data due to privacy regulations, fragmented data ownership, and regulatory constraints. However, these devices frequently receive third-party models or model updates through Firmware Over-The-Air (FOTA) mechanisms. This supply chain introduces severe security risks, especially backdoor attacks, where malicious triggers are embedded into models before deployment. As a result, IoT stakeholders must conduct network admission and pre-deployment auditing based solely on model weights, since clean datasets and trigger samples are not available for conventional analysis. Existing defense mechanisms that require input data or activation analysis are therefore impractical in IoT environments. To bridge this gap, we introduce a new paradigm for backdoor detection that inspects only the intrinsic properties of model weights. We propose Sequential Fingerprinting for Backdoor Detection (SFBD), a framework that treats a neural network as a structural sequence and constructs sequential fingerprints from model weights, enabling robust, data-free auditing before model deployment. By modeling cross-layer correlations, SFBD effectively identifies anomalies indicative of backdoors. In addition, SFBD is equipped with a few-shot learning mechanism, allowing it to rapidly adapt to novel OTA scenarios using only a small number of unlabeled candidate models. This makes SFBD an essential gatekeeper for secure IoT model updates. Comprehensive experiments across diverse backdoor attacks, datasets, model architectures, and a representative non-vision audio scenario show that SFBD achieves state-of-the-art performance in inference-data-free detection and strong cross-scenario generalization. Code is available at https://github.com/HaiyangYu-Lab/SFBD.
- 30 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP (15) , 1
SFBD: Backdoor Detection via Sequential Fingerprinting of Neural Networks for Securing the IoT Model Supply Chain
The rapid deployment of Deep Learning (DL) in the Internet of Things (IoT) increasingly relies on outsourcing model training to third parties or cloud servers due to the resource constraints of edge devices. In practical IoT scenarios, edge devices and gateways typically do not have access to original training data due to privacy regulations, fragmented data ownership, and regulatory constraints. However, these devices frequently receive third-party models or model updates through Firmware Over-The-Air (FOTA) mechanisms. This supply chain introduces severe security risks, especially backdoor attacks, where malicious triggers are embedded into models before deployment. As a result, IoT stakeholders must conduct network admission and pre-deployment auditing based solely on model weights, since clean datasets and trigger samples are not available for conventional analysis. Existing defense mechanisms that require input data or activation analysis are therefore impractical in IoT environments. To bridge this gap, we introduce a new paradigm for backdoor detection that inspects only the intrinsic properties of model weights. We propose Sequential Fingerprinting for Backdoor Detection (SFBD), a framework that treats a neural network as a structural sequence and constructs sequential fingerprints from model weights, enabling robust, data-free auditing before model deployment. By modeling cross-layer correlations, SFBD effectively identifies anomalies indicative of backdoors. In addition, SFBD is equipped with a few-shot learning mechanism, allowing it to rapidly adapt to novel OTA scenarios using only a small number of unlabeled candidate models. This makes SFBD an essential gatekeeper for secure IoT model updates. Comprehensive experiments across diverse backdoor attacks, datasets, model architectures, and a representative non-vision audio scenario show that SFBD achieves state-of-the-art performance in inference-data-free detection and strong cross-scenario generalization. Code is available at https://github.com/HaiyangYu-Lab/SFBD.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (18) , 1
STAPT: Spatio-Temporal Semantic Modeling for Advanced Persistent Threat Detection
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (18) , 1
STAPT: Spatio-Temporal Semantic Modeling for Advanced Persistent Threat Detection
- Open Access
- 20 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 2791-2802
BackdoorMBTI: A Backdoor Learning Multimodal Benchmark Tool Kit for Backdoor Defense Evaluation
- 20 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 2791-2802
Open AccessBackdoorMBTI: A Backdoor Learning Multimodal Benchmark Tool Kit for Backdoor Defense Evaluation