Reproduce “Leaking Circuit Secrets: Gradient Leakage Attacks on Graph Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rupesh Raj Karn, Johann Knechtel, Ozgur Sinanoglu

This study evaluates Gradient Leakage Attacks (GLAs) on Graph Neural Networks (GNNs) applied to circuit design and hardware security tasks, demonstrating that sensitive input features like gate types and Trojan characteristics can be reconstructed from gradients. It quantifies the inherent vulnerability of various GNN architectures and assesses the inconsistent efficacy and performance trade-offs of state-of-the-art defense mechanisms.

Key insights

  • Circuit-tailored GNNs are fundamentally vulnerable to GLAs, allowing reconstruction of sensitive input features like gate types or distinct Trojan properties from gradients.
  • The choice of GNN architecture significantly influences vulnerability to GLAs, with attention-based mechanisms (GAT) often exacerbating leakage in some tasks while injective aggregation (GIN) generally showed higher resilience.
  • Existing state-of-the-art defense mechanisms offer inconsistent mitigation against GLAs in the circuit domain; their effectiveness varies widely by GNN architecture, task, and even specific data class, sometimes leading to worsened leakage or significant performance degradation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.