HuggingGraph, the first work to systematically collect the LLM supply chain information and analyze it at scale.

• 📦 Systematically collect LLM supply chain data – We curate large-scale metadata connecting models and datasets across the LLM ecosystem.
• 🔗 Construct an LLM supply chain graph – We build a directed, heterogeneous graph that captures relationships between base models, fine-tuned models, adapters, quantized/merged variants, and datasets.
• 🔍 Perform rich graph analytics – We conduct forward and backward analyses to trace lineage, uncover dependencies, and derive insights into risks and evolution in the LLM supply chain.

SEGA-NET, the first framework to bridge LLM-guided semantic augmentation and robust low-resolution image classification.

• 🧠 Semantic-first prompt expansion – We use large language models to transform class labels into semantically diverse prompts, enabling controlled intra-class variation beyond pixel-level augmentation.
• 🎯 Projector-steered GAN synthesis – We introduce a lightweight text-to-latent projector that maps LLM semantics into a frozen class-conditional GAN, enabling efficient and label-consistent 32×32 generation.
• 🛡️ Security-aware data curation – We apply CLIP-based semantic filtering, SSIM de-duplication, and quality screening to retain only high-fidelity samples that improve accuracy and adversarial robustness.