An edge sensitivity based gradient attack on graph isomorphic networks for graph classification problems DOI Creative Commons
Srinitish Srinivasan,

Chandraumakantham OmKumar

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: April 29, 2025

Abstract Graph Neural Networks have gained popularity over the past few years. Their ability to model relationships between entities of same and different kind, represent molecules, flow etc. made them a go tool for researchers. However, owing abstract nature graphs, there exists no ideal transformation nodes edges in euclidean space. Moreover, GNNs are highly susceptible adversarial attacks. gradient based attack on latent space embeddings does not exist GNN literature. Such attacks, classified as white box tamper with representation graphs without creating any noticeable difference overall distribution. Developing testing models such attacks graph classification tasks would enable researchers understand develop stronger more robust systems. Further, tests literature been performed weaker, less representative neural network architectures. In order tackle these gaps literature, we propose developed from contrastive representations. strong base(victim) learning spectral spatial properties consideration isomorphic properties. We experimentally validate this 4 benchmark datasets molecular property prediction where our outperformed 75% all LLM-based On attacking proposed strategy, performance drops at an average 25% thereby clearing existent The code paper can be found https://github.com/Deceptrax123/An-edge-sensitivity-based-gradient-attack-on-GIN-for-inductive-problems

Language: Английский

Open-World Semi-Supervised Learning for fMRI Analysis to Diagnose Psychiatric Disease DOI Creative Commons

Chang Hu,

Yihong Dong, Shoubo Peng

et al.

Information, Journal Year: 2025, Volume and Issue: 16(3), P. 171 - 171

Published: Feb. 25, 2025

Due to the incomplete nature of cognitive testing data and human subjective biases, accurately diagnosing mental disease using functional magnetic resonance imaging (fMRI) poses a challenging task. In clinical diagnosis disorders, there often arises problem limited labeled due factors such as large volumes cumbersome labeling processes, leading emergence unlabeled with new classes, which can result in misdiagnosis. context graph-based disorder classification, open-world semi-supervised learning for node classification aims classify nodes into known classes or potentially presenting practical yet underexplored issue within graph community. To improve representation fMRI under low-label settings, we propose novel approach tailored analysis, termed Open-World Semi-Supervised Learning Analysis (OpenfMA). Specifically, employ spectral augmentation self-supervised dynamic concept contrastive achieve guided by pseudo-labels, construct hard positive sample pairs enhance network’s focus on potential pairs. Experiments conducted public datasets validate superior performance this method psychiatric domain.

Language: Английский

Citations

0
RRGMambaFormer: A hybrid Transformer-Mamba architecture for radiology report generation DOI
Hongzhao Li, Siwei Liu, Hui Wang

et al.

Expert Systems with Applications, Journal Year: 2025, Volume and Issue: unknown, P. 127419 - 127419

Published: April 1, 2025

Language: Английский

Citations

0
An edge sensitivity based gradient attack on graph isomorphic networks for graph classification problems DOI Creative Commons
Srinitish Srinivasan,

Chandraumakantham OmKumar

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: April 29, 2025

Abstract Graph Neural Networks have gained popularity over the past few years. Their ability to model relationships between entities of same and different kind, represent molecules, flow etc. made them a go tool for researchers. However, owing abstract nature graphs, there exists no ideal transformation nodes edges in euclidean space. Moreover, GNNs are highly susceptible adversarial attacks. gradient based attack on latent space embeddings does not exist GNN literature. Such attacks, classified as white box tamper with representation graphs without creating any noticeable difference overall distribution. Developing testing models such attacks graph classification tasks would enable researchers understand develop stronger more robust systems. Further, tests literature been performed weaker, less representative neural network architectures. In order tackle these gaps literature, we propose developed from contrastive representations. strong base(victim) learning spectral spatial properties consideration isomorphic properties. We experimentally validate this 4 benchmark datasets molecular property prediction where our outperformed 75% all LLM-based On attacking proposed strategy, performance drops at an average 25% thereby clearing existent The code paper can be found https://github.com/Deceptrax123/An-edge-sensitivity-based-gradient-attack-on-GIN-for-inductive-problems

Language: Английский

Citations

0