ObfuBench: A Benchmark Suite for HLS/RTL Hardware Obfuscation

ObfuBench is the first open benchmark suite designed for evaluating Machine Learning (ML), Deep Learning (DL), oracle-less, and oracle-guided attacks on HLS/RTL hardware obfuscation techniques.

---

Motivation

The globalization of the semiconductor supply chain has made hardware Intellectual Property (IP) increasingly vulnerable to threats such as IP piracy, reverse engineering, hardware Trojans, and malicious modifications. To address these threats, numerous High-Level Synthesis (HLS) and Register Transfer Level (RTL) obfuscation techniques have been proposed.

Recent advances in Machine Learning (ML) and Deep Learning (DL) have demonstrated the ability to compromise many hardware security mechanisms. However, unlike the gate-level domain—where public benchmarks such as Trust-Hub are available—no standardized benchmark exists for evaluating ML/DL attacks on HLS/RTL obfuscation techniques.

As a result, researchers often generate their own datasets using different methodologies, making fair comparison, reproducibility, and benchmarking difficult.

ObfuBench addresses this gap by providing a unified benchmark suite containing datasets generated from multiple state-of-the-art HLS/RTL obfuscation techniques using a standardized generation methodology.

---

Features

• Standardized benchmark for HLS/RTL hardware obfuscation

• Multiple state-of-the-art obfuscation techniques

• Large-scale datasets for ML/DL training and evaluation

• Four balanced key patterns

• Four obfuscation coverage levels

• Functionally equivalent but structurally unique hardware designs

• Suitable for:

  • Machine Learning attacks
  • Deep Learning attacks
  • Graph Neural Network (GNN) attacks
  • Oracle-less attacks
  • Oracle-guided attacks
  • Hardware security benchmarking

---

Benchmark Configuration

Each obfuscation technique is generated using the same configuration.

Key Patterns

Four balanced secret keys are used throughout the benchmark:

01010101...
10101010...
01100110...
10011001...

Each key contains an equal number of 0s and 1s, eliminating bias during ML/DL model training.

Obfuscation Coverage

Each benchmark is generated with four protection levels:

• 25%
• 50%
• 75%
• 100%

Dataset Size

For every:

• Obfuscation technique
• Key pattern
• Obfuscation level

the benchmark contains:

5,000 structurally unique but functionally equivalent hardware designs.

---

Included Obfuscation Techniques

PROTECTS

Reference

PROTECTS: Progressive RTL Obfuscation with Threshold Control Technique During Architectural Synthesis.

• 4 balanced key patterns
• 25%, 50%, 75%, and 100% obfuscation
• 5,000 variants per key pattern

---

KOIL

Reference

High-Level Synthesis of Key-Obfuscated RTL IP with Design Lockout and Camouflaging.

• 4 balanced key patterns
• 25%, 50%, 75%, and 100% obfuscation
• 5,000 variants per key pattern

---

ILP

Reference

Low Cost Functional Obfuscation of Reusable IP Cores Used in Consumer Electronics Hardware Through Robust Locking.

• 4 balanced key patterns
• 25%, 50%, 75%, and 100% obfuscation
• 5,000 variants per key pattern

---

ASSURE

Reference

ASSURE: RTL Locking Against an Untrusted Foundry. IEEE Transactions on Very Large Scale Integration (VLSI) Systems.

• 4 balanced key patterns
• 25%, 50%, 75%, and 100% obfuscation
• 5,000 variants per key pattern

---

Repository Structure

ObfuBench/
├── PROTECTS/
├── KOIL/
├── ILP/
├── ASSURE/
└── README.md

Each technique contains datasets generated using:

• 4 balanced key patterns
• 4 obfuscation coverage levels
• 5,000 structurally unique hardware variants per configuration

---

Applications

ObfuBench can be used for:

• Benchmarking HLS/RTL obfuscation techniques
• Machine Learning attacks
• Deep Learning attacks
• Oracle-less attacks
• Oracle-guided attacks
• Graph Neural Network research
• Hardware security evaluation
• Reproducible academic research

---

Citation

If you use ObfuBench in your research, please cite our accompanying paper.