H2-GAT: Hierarchical Heterogeneous Graph Attention Networks for AI Model Safety Verification

Abstract

The proliferation of AI models from open-source platforms has introduced significant security risks, as these models may contain hidden vulnerabilities that compromise application security. Existing static analysis tools rely on fixed rule sets that fail to generalise to novel AI-specific threat patterns, while LLM-based approaches lack interpretability and require orders of magnitude more parameters. We propose Hierarchical Heterogeneous Graph Attention Networks (H2-GAT), which combines comprehensive program analysis with optimised graph neural network architectures for automated vulnerability detection. Our method constructs heterogeneous graphs with six essential edge types capturing Security-Enriched Abstract Syntax Tree (SE-AST), Control Flow Graph (CFG), Data Flow Graph (DFG), and taint flow semantics, then applies hierarchical attention mechanisms for graph-level vulnerability detection. H2-GAT achieves a peak F1-score of 98.2% with 96.4% precision, using only 226, 387 parameters—approximately 550× fewer than CodeBERT. This enables efficient CPU deployment while providing auditable, graph-traceable reasoning unavailable in opaque LLM-based detectors. The system successfully identifies SQL injection, command injection, code injection, model poisoning, and backdoor-insertion patterns while maintaining computational efficiency through semantic edge-type consolidation.