AttnCall: Refining Indirect Call Targets in Binaries with Attention

Accurate Control Flow Graphs are crucial for effective binary program analysis, while solving indirect function call targets is its major challenge. Existing static analysis methods heavily rely on domain-specific patterns, resulting in an abundance of false positive edges due to limited expert knowledge. Concurrently, learning-based approaches often depend on heuristic analysis during the code representation stage, which prevents the model from fully comprehending program semantics. To address these limitations, this paper presents AttnCall, a novel neural network learning framework that leverages the attention mechanism to automatically learn the matching relationship between function callsites and callees’ context semantics. AttnCall refines the identification of indirect call targets through the learned matching patterns, eliminating the drawbacks of existing techniques. Additionally, we propose an end-to-end code representation scheme that effectively embeds the semantics of callsites and callees without relying on heuristic rules. The evaluation of AttnCall focuses on the task of predicting indirect function call targets. The results demonstrate that AttnCall surpasses state-of-the-art approaches, achieving 31.4% higher precision and 5% higher recall. Moreover, AttnCall enhances model interpretability, allowing for a better understanding of the underlying analysis process.