Full paper: www.plai.ifi.lmu.de/publications...

BLens achieves a 42% increase in F1 score on unseen projects, also improving RougeL, Bleu, and VarCLR for function names. Performance metrics tell only part of the story, however, and the paper shows how even seemingly mispredicted functions can provide valuable information to a reverse engineer.

Our model architecture has two main components, COMBO for pretraining and LORD for finetuning and inference. COMBO uses a contrastive captioning loss (inspired by CoCa), whereas LORD uses an MLM objective to optimize the generation of function names while reducing false positives.

Our intuition is that naming binary functions is similar to image captioning, with function names and code as two modalities for the same concept. Inspired by the multimodal image-text model GIT, we embed functions into patches using three pre-trained function embeddings (DEXTER, CLAP, PalmTree).

Binary reversing is hard. One approach to helping is to generate plausible names for binary functions using machine learning models. But state-of-the-art models based on the idea of translating assembly to English struggle with generalizing across projects and may generate misleading names.

Apart from an exciting main program with keynote speakers Shweta Shinde @shwetashinde.bsky.social and Bart Preneel, we will have nine workshops ranging from IoT security to post-quantum cryptography, a poster reception, and a conference dinner at the iconic Löwenbräukeller.