REVDECODE: Enhancing Binary Function Matching with Context-Aware Graph...
summary
REVDECODE converts function-matcher outputs into directed layered graphs, runs a Viterbi-inspired forward/backward pass to find maximum-weight paths, and re-ranks matches using contextual edges and confidence/library/adjacency weights. GPU-optimized variants partition traversal for parallelism and a segment-based algorithm trades ~0.003 loss in accuracy for large speedups. Evaluated on 11 billion and 5.8 billion function-match pairs across two datasets, it improves 56.3%–98.8% of function rankings for four existing matchers by promoting relevant but structurally dissimilar versions and resolving ambiguous duplicates.
AI-written summary · about
chapters (10)
- 0:00Function matching challenges
- 1:37Current function matching limits
- 3:23Introducing Revcode's relevance decoding
- 4:48Revcode’s graph-based framework
- 7:36Evaluation methodology
- 9:01Improving match rankings
- 10:27Component analysis
- 11:11Parallelizing graph traversal
- 12:38Parallel algorithm results
- 12:59Summary of Revcode
related talks
41mfn_fuzzy: Fast Multiple Binary Diffing Triage With IDA
16mCodeExtract: Enhancing Binary Code Similarity Detection with Code Extraction Techniques
30mBoosting Vulnerability Intelligence: How Accurate CWE Mappings Transform ML Model Performance
30mEnhancing CTI Processes with Code Search Technology
22mA Scalable, Ensemble Approach for Building and Visualizing Deep Code-Sharing Networks
21mAutomatic Recovery of Fine-grained Compiler Artifacts at the Binary Level
14mNeural Reverse Engineering of Stripped Binaries using Augmented Control Flow Graphs
31mAn AI Approach to Malware Similarity Analysis: Mapping the Malware Genome With a Deep Neural Network