OPTango: Multi-central Representation Learning against Innumerable Compiler Optimization for Binary Diffing

Binary diffing, which quantitatively measures the difference between given binaries, has been broadly used in critical security areas. Previous studies have been tackling the challenge of default compiler optimization, as it can affect binary representation but overlooked the exploration of non-default optimization settings, which can also significantly affect the accuracy of diffing. Recent research indicates a growing trend of compiling applications with non-default optimization settings to magnify binary code discrepancies, enabling them to evade detection by binary diffing tools. This paper takes the first step to systematically studying the resistance of compiler optimization (including default and non-default optimization settings) on binary diffing tasks. To this end, we construct a diverse and unique dataset, OPTBinary, with 3.6 million functions compiled from 514 optimization settings. Then, we propose OPTango, an innovative transformer-based multi-central representation learning approach, exploring the solution to build a compiler optimization-agnostic binary diffing tool. We conduct extensive experiments and benchmark OPTango with state-of-the-art binary diffing approaches. Evaluation results show that OPTango is more robust and significantly outperforms existing methods against both default and non-default compiler optimization.