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. Our study 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 combined, 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, with minimal area and power overheads.
Introduction: The Vulnerability of Globalized IC Design
In the modern semiconductor industry, the "fabless" model dominates, where design houses outsource manufacturing to offshore foundries. This globalization introduces significant security risks, including IP theft and reverse engineering by untrusted foundries. Logic locking has emerged as a primary defense, where designers insert "key-gates" (e.g., XOR/XNOR) into the netlist. The circuit only functions correctly when the correct key is loaded into secure memory.
However, the security of logic locking is threatened by "oracle-less" attacks like SAIL (Structural Analysis for Logic Locking). Unlike traditional attacks that require a functional chip (an "oracle"), SAIL uses machine learning to analyze structural footprints left by synthesis tools. By recognizing how these tools transform the circuit, ML models can predict the correct keys with high accuracy.
The SAIL Attack Mechanism
The SAIL attack exploits the deterministic nature of logic synthesis. When a locked netlist is synthesized, the tool optimizes the logic around key-gates, creating recognizable patterns. The attack pipeline typically involves extracting subgraphs from a post-synthesis netlist and using ML models to reconstruct the original pre-synthesis structure, eventually leading to key recovery.
Methods: Thwarting ML Attacks with UNSAIL
UNSAIL is a proactive defense designed to "poison" the training data of ML-based attacks. The core strategy is adversarial data injection, which breaks the clear classification of circuit subgraphs into "Changed" (transformed by synthesis) and "Unchanged" categories.
The UNSAIL methodology involves several key steps:
- Partial Traditional Locking: Half of the key-bits are locked using standard methods, followed by synthesis to observe structural changes.
- Subgraph Analysis: The design is analyzed to identify which subgraphs were altered and which remained stable.
- Strategic Injection: For the remaining key-bits, UNSAIL injects subgraphs into the post-synthesis netlist that mimic the "opposite" label. For instance, it adds a key-gate to a structure that appears "Unchanged," confusing the ML model's feature space.
- Compound Locking: UNSAIL utilizes a variety of key-gate types (e.g., MUXes, AND/OR gates) to diversify the circuit's structural vocabulary, further disrupting the patterns relied upon by attackers.
Results and Evaluation
The effectiveness of UNSAIL was evaluated using industry-standard benchmarks, including ISCAS-85, ITC-99, and the million-gate ORPSoC design.
- Attack Degradation: UNSAIL significantly reduced the accuracy of SAIL’s change-prediction and reconstruction models by an average of 20.13 and 17 percentage points, respectively.
- General Resilience: The technique also proved effective against other oracle-less attacks, reducing SWEEP attack accuracy by 15 pp and limiting redundancy attacks to 38% key recovery.
- Overhead: On the ORPSoC GPS module, UNSAIL incurred minimal area and power overheads of 0.26% and 0.61%, respectively, demonstrating its feasibility for large-scale designs.
Conclusion
UNSAIL represents a shift from reactive to proactive hardware security. By targeting the training phase of ML attacks, it turns the synthesis process—once an attacker's advantage—into a tool for obfuscation. While primarily effective against structural attacks, future work involves pairing UNSAIL with functional defenses (like SAT-resilient locking) to provide a comprehensive multi-layer protection scheme for modern ICs.
Summary updated