Titan: Security Analysis of Large-Scale Hardware Obfuscation Using Graph Neural Networks

Hardware obfuscation is a prominent design-for-trust solution that thwarts intellectual property (IP) piracy and reverse-engineering of integrated circuits (ICs). Researchers have proposed several large-scale obfuscation techniques that achieve high output corruption&#x2014;thus offering resilience against seminal attacks along with acceptable power, performance, and area overheads. However, the research community has primarily evaluated hardware obfuscation on relatively small scales of obfuscation (<italic>i.e</italic>., a fixed number of obfuscated components). Moreover, prior art caters toward specific schemes based either on gate obfuscation or interconnect obfuscation, <italic>i.e</italic>., two prominent types of hardware obfuscation. The former shortcoming suggests focusing on large-scale obfuscation schemes, and the latter suggests the need for a holistic assessment framework. In this work, we propose Titan, a holistic framework considering large-scale gate and interconnect obfuscation schemes. More specifically, we propose a graph neural network (GNN)-based attack framework that is trained to exploit structural and functional properties of any secured circuit to recover its obfuscated components.We evaluate Titan on various obfuscation schemes, considering selected ITC-99 benchmarks with up to 50% obfuscation scale, <italic>i.e</italic>., up to 21,326 obfuscated components. We observe a substantial information leakage through structural and functional properties of secured designs even for large-scale obfuscation. We quantify the information leakage in two ways: first, an average reduction of Hamming distance (HD, a well-established metric for attack evaluation) by 23.27 and 16.19 percentage points over the baseline of random guessing for gate and interconnect obfuscation, respectively; second, an average recovery of 63.40% and 77.94% of obfuscated components for gate and interconnect obfuscation, respectively. Importantly, these results are superior to six state-of-the-art attacks. We will open-source our framework and associated artifacts to enable reproducibility and foster future work.