Research map / Detection and segmentation / CNN features
BDLF-Qwen3: Enhanced Cross-Architecture Binary Function Similarity Detection Through Binary Dynamic Layer Fusion
Yuanda Wang, Ji Zhou, Xinhui Han, Chao Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic input-dependent layer weights (ablation) | BDLF-Qwen3 · Cross-Bin poolsize 1,000 | Recall@10: 0.59 -> 0.67 | Improved |
| use dynamic weights instead of learned static weights (ablation) | BDLF-Qwen3 · Cross-Bin poolsize 1,000 | Recall@10: 0.60 -> 0.67 | Improved |
| add XRef-enhanced information to Cross-Bin (ablation) | BDLF-Qwen3 · Cross-Bin poolsize 10,000 | Recall@10: 0.335 -> 0.378 | Improved |
| add stage 1 partial fine-tuning before BDLF stage (ablation) | Qwen3-Embedding BDLF · Cross-Bin poolsize 1,000 | NDCG@10: 0.50 -> 0.65 | Improved |
| add BDLF stage 2 after stage 1 fine-tuning (ablation) | Qwen3-Embedding · Cross-Bin poolsize 1,000 | NDCG@10: 0.58 -> 0.65 | Improved |
| add BDLF stage 2 after stage 1 fine-tuning (ablation) | UnixCoder · Cross-Bin poolsize 1,000 | NDCG@10: 0.44 -> 0.51 | Improved |
| add BDLF stage 2 after stage 1 fine-tuning (ablation) | GraphCodeBERT · Cross-Bin poolsize 1,000 | NDCG@10: 0.31 -> 0.40 | Improved |
| BDLF-Qwen3 versus UnixCoder on unseen RISC-V | zero-shot cross-arch search · RISC-V (poolsize 1000) | Recall@10: 0.38 -> 0.58 | Improved |
| BDLF-Qwen3 versus UnixCoder on vulnerability discovery | cross-arch CWE search · 4 CWE types | Recall@10 (avg): 0.705 -> 0.945 | Improved |
| scale pool size from small to 10,000 (ablation) | BDLF-Qwen3 · Cross-Bin | Recall@10: 0.905 -> 0.383 | Hurt |
Closest papers by meaning
- Is Function Similarity Over-Engineered? Building a Benchmark (NeurIPS 2024 · similarity 0.84)
- AdaCCD: Adaptive Semantic Contrasts Discovery Based Cross Lingual Adaptation for Code Clone Detection (AAAI 2024 · similarity 0.83)
- Nova: Generative Language Models for Assembly Code with Hierarchical Attention and Contrastive Learning (ICLR 2025 · similarity 0.81)
- CODE REPRESENTATION LEARNING AT SCALE (ICLR 2024 · similarity 0.81)
- ASAP: Assembly-Source Aligned Pseudocode Refinement For Binary Decompilation (NeurIPS 2026 · similarity 0.80)
- Source Code Foundation Models are Transferable Binary Analysis Knowledge Bases (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
