MicroSim reports high binary-similarity scores on TREX microcode graphs
Original title: MicroSim: Contrastive Graph Neural Networks for Binary Code Similarity Detection
On a class-disjoint TREX split, the authors report 0.9972 ROC-AUC and 0.9762 MCC, above retrained Gemini, SAFE, and jTrans.
MicroSim represents each binary function as a graph from IDA Pro Hex-Rays microcode and compares graphs with a contrastive Siamese graph neural network. On the TREX cross-optimization benchmark, 849,072 function-level graphs span optimization levels -O0 through -O3 under a class-disjoint split. The authors report 0.9972 ROC-AUC and 0.9762 MCC, above retrained Gemini, SAFE, jTrans, and a bag-of-words baseline. They also compare ranking fixes and loss functions, finding InfoNCE stronger for thresholded classification and retrieval.
Unconfirmed: Scores are author-reported; no independent replication is stated.
- The test split shares no function key with training.
- Gemini, SAFE, and jTrans were retrained on the same split.
- CSLS raises GraphGPS MRR from 0.2542 to 0.2559.
- Reported GraphGPS recall is 0.1639 at 1 and 0.4561 at 10.
- InfoNCE beats cosine-embedding loss on classification and retrieval.
- Vendors
- Hex-Rays
From the original article
1 min read at arxiv.orgWe present MicroSim, a system for binary code similarity detection that represents each function as a graph built from IDA Pro Hex-Rays microcode and compares such graphs with a contrastive Siamese graph neural network (GNN).
By Moussa Moussaoui · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.