My research asks how we can make AI systems traceable, auditable, and secure as they are developed, adapted, and deployed. I focus on AI supply-chain security, model provenance, and vulnerability propagation across base models and their derivatives.
Modern AI is rarely built from scratch. Models are fine-tuned, quantized, merged, extended with adapters, and deployed inside autonomous agents. I use graph-based methods, static analysis, and large-scale empirical evaluation to understand these relationships and how security risks travel through them. My work also explores agent security and trustworthy, resource-efficient edge AI.
Open model hubs such as Hugging Face now host millions of models. Each one may derive from a base model through fine-tuning, adapters, quantization, or merging, and depend on datasets and code from many sources. These relationships form a large and largely undocumented AI supply chain. Lineage declarations are often missing, incomplete, or inconsistent, which makes it hard to answer basic security questions: Where did this model come from? What does it inherit? Who is affected when an upstream model is found to be unsafe?
My research models this ecosystem as a graph and develops methods to characterize, audit, and recover model provenance at scale. Selected projects are outlined below.
HuggingGraph: Understanding the Supply Chain of LLM Ecosystem
Built a heterogeneous graph of the Hugging Face LLM ecosystem with 402,654 entities and 462,524 relationships, capturing model derivation, fine-tuning, adapters, quantization, merging, and dataset dependencies.
Analyzed ecosystem structure, connectivity, and provenance gaps to enable graph-based AI supply-chain analysis.
Released public graph datasets, documentation, and a companion repository to support reproducibility.
ProvCode: Static Auditing of Provenance Carryover for Hugging Face Models
Under review, ICSE 2027
Developed static-analysis methods to check whether model repositories and their source code preserve upstream provenance claims.
Examined inconsistencies between declared and code-derived lineage across roughly 75,000 Hugging Face models.
ChainInfer: Graph-Based Model Provenance Inference
Under review, WWW
Investigates recovery of missing model-to-model lineage relationships using graph topology, repository metadata, and learned representations.
Aims to reconstruct hidden upstream dependencies that model cards and metadata fail to declare.
Large-Scale Model Lineage Analysis
Ongoing
Extended data collection and dependency extraction to about three million Hugging Face model records, studying missing, incomplete, and nonstandard lineage declarations across the broader ecosystem.
02
Vulnerability Propagation Across Model Derivatives
If a base model behaves unsafely, do its descendants inherit that behavior? Supply-chain security depends on understanding how vulnerabilities propagate through adaptation. This line of work runs controlled, large-scale experiments to measure whether security weaknesses persist, weaken, or amplify as models are fine-tuned, adapted, and compressed.
VulChain: AI Supply-Chain Vulnerability Propagation
Manuscript Revision in Progress
Designed controlled empirical evaluations to test whether unsafe behaviors, including package hallucination and insecure URL generation, persist through fine-tuning, adapter-based adaptation, and quantization.
Measured vulnerability persistence using matched prompts and validated detectors across base models and their derivatives.
ChronoResolve: Retrospective Artifact Resolution for Model Lineage
Research in Progress
Investigates the identification of historical upstream artifacts a model was derived from, using file hashes, configurations, weight signatures, and temporal evidence.
Aims to reduce lineage ambiguity to support reproducible provenance and security auditing.
03
Agentic and Multi-Agent AI Security
LLMs are increasingly deployed as autonomous agents that call tools, process multimodal inputs, and collaborate with other agents. This creates new attack surfaces: a single injected prompt can cross trust boundaries, spread between agents, and trigger unsafe real-world actions. My research builds systematic tools to expose these failures and defend against them at runtime.
AgentFuzz: Trace-Level Fuzzing for Prompt Spillover in LLM Multi-Agent Systems
Developed a trace-level fuzzing framework to study how adversarial prompts spill over across interacting LLM agents.
Evaluated attack propagation across communication topologies, injection channels, and verification configurations to characterize security failures in multi-agent workflows.
DiscoAct: Securing Autonomous Multimodal LLM Agents Against Cross-Modal Trust Flattening
Developed a model-agnostic runtime verification framework that combines symbolic policy validation, cross-modal consistency checking, and knowledge–environment alignment.
Evaluated it on 240 adversarial scenarios across three LLM backbones, preventing 94.2% of unsafe actions.
04
Trustworthy and Resource-Efficient Edge AI
Deploying AI on edge devices means working with tight compute budgets and often low-quality inputs. This work develops lightweight, semantic-guided learning methods that keep visual recognition robust and reliable under these constraints.
LiteAugNet: Lightweight Semantic-Guided Augmentation for Edge-Level Image Classification
A lightweight semantic-guided augmentation network that improves low-resolution image classification while respecting the compute limits of edge environments.
SEGA-Net: LLM-Guided Semantic-Enhanced GAN Augmentation for Low-Resolution Image Classification
