Golden-Reference Model Fuzzing
Overview
Golden-reference model fuzzing is a technique in hardware security testing that employs a two-stage fuzzing strategy. The approach decouples test refinement from vulnerability exploration by using a fast golden-reference model of the device under test. This separation enables efficient and semantically aware hardware fuzzing, achieving high coverage and discovering new high-severity vulnerabilities on target processors.
How It Works
The technique operates in two stages:
- Test refinement — A fast golden-reference model of the device is used to rapidly iterate and refine test inputs, providing quick feedback on behavioral correctness.
- Vulnerability exploration — Refined tests are then applied against the real (or more detailed) implementation to uncover subtle bugs and security vulnerabilities.
By decoupling these phases, golden-reference model fuzzing avoids the cost of running expensive target evaluations during every mutation while still maintaining semantic awareness of the device's expected behavior.
Demonstrated Application
Golden-reference model fuzzing was demonstrated on RISC-V processors as part of the GoldenFuzz framework, where it achieved:
- High coverage of processor behavior
- Discovery of new high-severity vulnerabilities
Context
The technique was presented at NDSS 2026 in the "Fuzzing – Shake the Silicon" session, alongside related hardware-fuzzing research such as ReFuzz (processor fuzzing via contextual bandits reusing historic tests) and Fuzzilicon (post-silicon microcode-guided x86 CPU fuzzing).
Related Implementations and Publications
- GoldenFuzz (Tool) — implements golden-reference model fuzzing.
- GoldenFuzz: Generative Golden Reference Hardware Fuzzing (Paper) — introduces and uses the technique.