Abstract:
Logic locking technology aims to protect hardware intellectual property against reverse engineering and unauthorized access. Although attacks leveraging graph neural networks (GNNs) have demonstrated notable success in compromising logic locking, their performance remains heavily contingent upon high-quality training data. However, in practical scenarios, the prohibitive cost of reverse engineering makes acquiring sufficient locked circuit netlists extremely challenging, severely constraining the improvement of attack accuracy. To overcome these limitations, we propose a novel hybrid two-phase attack method named EAGLE that combines GNN capabilities with advanced combinatorial optimization techniques to substantially enhance overall attack efficacy. In the initial phase, a fine-tuned large language model (LLM) is employed to synthesize high-fidelity training data, which not only refines GNN prediction accuracy but also generates reliable partial key outputs. Subsequently, the second phase utilizes a binary differential evolution algorithm to iteratively optimize the remaining unknown key bits, ultimately recovering the complete decryption keys. Furthermore, ensemble learning strategies are integrated into this second phase to maximize robustness and precision. Experimental results demonstrate that in the oracle-less scenario, the first phase of EAGLE achieves high-confidence predictions with over 93.6% accuracy for 60.11% of the key bits, significantly outperforming existing state-of-the-art GNN-based attack methods. Furthermore, in the oracle-guided scenario, by integrating binary differential evolution and ensemble learning mechanisms in the second phase, the overall average accuracy of the complete two-phase EAGLE method is further boosted to 95.09%.