BinFormer: Function Similarity Detection through Function Name Prediction - Reproducibility Package

Reproducibility Package for "BinFormer: Function Similarity Detection through Function Name Prediction" (MSc Praktikum Binary Analysis with AI, LMU, 2026) This package accompanies the project BinFormer: Function Similarity Detection through Function Name Prediction, which investigates whether function name prediction can serve as a proxy task for binary function similarity detection in assembly code. The project introduces and evaluates the BinFormer architecture, a transformer-based approach designed to learn representations of binary functions from assembly instructions. The package includes: Model architectures and training pipelines for the BinFormer encoder, similarity classification model, and encoder–decoder function name prediction model Dataset preparation and analysis scripts, including tools for processing binary datasets and extracting assembly-level features using binary analysis frameworks Experimental pipelines for evaluating the proxy similarity task, where predicted function names are used to infer similarity between binary functions This reproducibility package enables full replication of the experiments presented in the report, including training the baseline masked language model, fine-tuning similarity models, and evaluating the effectiveness of function name prediction as a proxy for binary similarity. It provides a foundation for further research in machine learning–based binary analysis and reverse engineering.