Beyond Sparse Supervision: Diffusion-Guided Learning for Few-Shot Graph Fraud Detection

Researchers have proposed two innovative methods to enhance safety and security in different domains: ADC-GNN for few-shot graph fraud detection and a radar-guided camera verification method for Automatic Emergency Braking (AEB) systems.

Graph-based fraud detection is crucial for protecting large-scale transaction systems, where undetected anomalies can result in significant financial losses and security risks^[1]. Real-world fraud graphs present challenges such as sparse and imbalanced supervision, and representation dilution. To address these issues, researchers have developed ADC-GNN, a framework that combines diffusion-guided feature augmentation, contrastive representation learning, and multi-hop spectral attention. ADC-GNN has shown consistent improvements over baseline models on public benchmarks. In a separate development, a radar-guided camera verification method has been proposed for AEB systems, which are widely used in vehicles. Radar provides reliable range and velocity measurements, while cameras offer visual confirmation of objects^[2]. The proposed method, which requires no training data or GPU acceleration, was integrated into a complete radar-camera fusion AEB system and evaluated on a real instrumented vehicle across 72 driving sessions and 131,603 camera frames^[2].