UNSAIL: Thwarting Oracle-Less Machine Learning Attacks on Logic Locking

Lilas Alrahis, Satwik Patnaik, Johann Knechtel, Hani Saleh, Baker Mohammad, Mahmoud Al‐Qutayri, Ozgur Sinanoglu

Abstract

Logic locking aims to protect the intellectual property (IP) of integrated circuit (IC) designs throughout the globalized supply chain. The SAIL attack, based on tailored machine learning (ML) models, circumvents combinational logic locking with high accuracy and is amongst the most potent attacks as it does not require a functional IC acting as an oracle. In this work, we propose UNSAIL, a logic locking technique that inserts key-gate structures with the specific aim to confuse ML models like those used in SAIL. More specifically, UNSAIL serves to prevent attacks seeking to resolve the structural transformations of synthesis-induced obfuscation, which is an essential step for logic locking. Our approach is generic; it can protect any local structure of key-gates against such ML-based attacks in an oracle-less setting. We develop a reference implementation for the SAIL attack and launch it on both traditionally locked and UNSAIL-locked designs. For SAIL, two ML models have been proposed (which we implement accordingly), namely a change-prediction model and a reconstruction model; the change-prediction model is used to determine which key-gate structures to restore using the reconstruction model. Our study on benchmarks ranging from the ISCAS-85 and ITC-99 suites to the OpenRISC Reference Platform System-on-Chip (ORPSoC) confirms that UNSAIL degrades the accuracy of the change-prediction model and the reconstruction model by an average of 20.13 and 17 percentage points (pp), respectively. When the aforementioned models are combined, which is the most powerful scenario for SAIL, UNSAIL reduces the attack accuracy of SAIL by an average of 11pp. We further demonstrate that UNSAIL thwarts other oracle-less attacks, i.e., SWEEP and the redundancy attack, indicating the generic nature and strength of our approach. Detailed layout-level evaluations illustrate that UNSAIL incurs minimal area and power overheads of 0.26% and 0.61%, respectively, on the million-gate ORPSoC design.

Introduction

The globalization of the semiconductor supply chain has introduced significant security risks for fabless design houses. Outsourcing fabrication to offshore foundries exposes sensitive Intellectual Property (IP) to potential theft, illegal overproduction, and reverse engineering. Logic locking has emerged as a primary defense, where a design is "locked" by inserting additional key-gates. The resulting circuit only functions correctly when the correct secret key is applied. However, the security of logic locking is constantly challenged by increasingly sophisticated attacks.

The logic locking and attack landscape, highlighting the distinction between oracle-less and oracle-guided attacks.

Traditionally, attacks were "oracle-guided," requiring access to a functional, activated chip to query inputs and observe outputs. More recently, "oracle-less" attacks have gained prominence. These attacks, such as SAIL (Structural Analysis for Logic Locking), rely solely on the gate-level netlist. By leveraging machine learning (ML), these attacks can identify patterns left by synthesis tools to reverse-engineer the original logic, bypassing the need for a working oracle.

The Threat of Oracle-Less Attacks: SAIL

The potency of the SAIL attack lies in its ability to learn how logic synthesis tools transform circuit structures. When key-gates are inserted, the design undergoes re-synthesis to optimize and further obfuscate the logic. Because synthesis tools are deterministic, they transform specific subgraphs in predictable ways. SAIL uses ML models—specifically a change-prediction model and a reconstruction model—to identify these transformations. The change-prediction model flags subgraphs that have likely been altered, and the reconstruction model attempts to restore them to their original, pre-synthesis state. This structural recovery effectively "unlocks" the design without ever needing the secret key.

UNSAIL: Camouflaging the Netlist

To counter these "smart" attacks, we propose UNSAIL. The core philosophy of UNSAIL is to break the deterministic link between synthesis and structural transformation that ML models exploit. Instead of trying to hide the gates better, UNSAIL intentionally injects "noise" into the netlist that mimics the "signal" of synthesis transformations.

The UNSAIL methodology, showing the integration of change prediction and reconstruction models to thwart key-guessing.

The UNSAIL workflow involves a strategic two-stage locking process. First, a portion of the design is locked traditionally and synthesized. The resulting structural changes are analyzed and stored in a dictionary. For the remaining key-bits, UNSAIL manually inserts structures that are indistinguishable from the synthesized ones, but which do not represent the same underlying logic transformations. This creates a "structural overlap" where the same local subgraph could represent multiple different original states, effectively camouflaging the real logic. By poisoning the attacker's feature space, UNSAIL forces the ML models to make incorrect predictions.

Experimental Results

We evaluated UNSAIL against a reference implementation of the SAIL attack across several benchmark suites, including ISCAS-85, ITC-99, and the million-gate ORPSoC design. Our results demonstrate that UNSAIL significantly degrades attack performance. The accuracy of the change-prediction model dropped by an average of 20.13 percentage points (pp), while the reconstruction model's accuracy fell by 17 pp. In the most comprehensive attack scenario, the overall accuracy of SAIL was reduced by 11 pp, often approaching the 50% mark of random guessing.

Beyond SAIL, UNSAIL proved effective against other oracle-less threats like SWEEP and redundancy-based attacks, highlighting its generic defensive capabilities. Importantly, these security gains come at a very low cost. Layout-level evaluations on the ORPSoC design showed minimal overheads, with area and power increasing by only 0.26% and 0.61%, respectively.

Conclusion

As machine learning becomes a standard tool for hardware reverse engineering, defenses must evolve to address structural vulnerabilities. UNSAIL provides a robust, low-overhead solution that thwarts ML-based attacks by introducing structural ambiguity. By camouflaging the fingerprints of synthesis, UNSAIL ensures that logic locking remains a viable defense for protecting IP in a globalized manufacturing environment.

Summary updated

Abstract

Logic locking aims to protect the intellectual property (IP) of integrated circuit (IC) designs throughout the globalized supply chain. The SAIL attack, based on tailored machine learning (ML) models, circumvents combinational logic locking with high accuracy and is amongst the most potent attacks as it does not require a functional IC acting as an oracle. In this work, we propose UNSAIL, a logic locking technique that inserts key-gate structures with the specific aim to confuse ML models like those used in SAIL. More specifically, UNSAIL serves to prevent attacks seeking to resolve the structural transformations of synthesis-induced obfuscation, which is an essential step for logic locking. Our approach is generic; it can protect any local structure of key-gates against such ML-based attacks in an oracle-less setting. We develop a reference implementation for the SAIL attack and launch it on both traditionally locked and UNSAIL-locked designs. For SAIL, two ML models have been proposed (which we implement accordingly), namely a change-prediction model and a reconstruction model; the change-prediction model is used to determine which key-gate structures to restore using the reconstruction model. Our study on benchmarks ranging from the ISCAS-85 and ITC-99 suites to the OpenRISC Reference Platform System-on-Chip (ORPSoC) confirms that UNSAIL degrades the accuracy of the change-prediction model and the reconstruction model by an average of 20.13 and 17 percentage points (pp), respectively. When the aforementioned models are combined, which is the most powerful scenario for SAIL, UNSAIL reduces the attack accuracy of SAIL by an average of 11pp. We further demonstrate that UNSAIL thwarts other oracle-less attacks, i.e., SWEEP and the redundancy attack, indicating the generic nature and strength of our approach. Detailed layout-level evaluations illustrate that UNSAIL incurs minimal area and power overheads of 0.26% and 0.61%, respectively, on the million-gate ORPSoC design.
BibTeX
@article{alrahis2021i,
  author = {Lilas Alrahis and Satwik Patnaik and Johann Knechtel and Hani Saleh and Baker Mohammad and Mahmoud Al‐Qutayri and Ozgur Sinanoglu},
  title = {{<i>UNSAIL</i>: Thwarting Oracle-Less Machine Learning Attacks on Logic Locking}},
  journal = {IEEE Transactions on Information Forensics and Security},
  year = {2021},
  doi = {10.1109/tifs.2021.3057576},
}
Figure