BERT-based model for identifying function and instruction boundaries in stripped x86 binary files.
Training
- Dataset: BigQuery C dataset, ~140k binaries, compiled with GCC at
-O1
- Epochs: 5
- Batch size: 256
- Learning rate: 1e-4
- Class weight: 5
- Init loss weight: 1
Classification Report
--- Function Boundary Classification Report ---
Class Distribution:
O (non-boundary): 78,124,536 (99.26%)
B-FUNC: 298,996 (0.38%)
E-FUNC: 281,198 (0.36%)
precision recall f1-score support
O 1.00 1.00 1.00 78124536
B-FUNC 0.95 0.99 0.97 298996
E-FUNC 0.95 0.99 0.97 281198
accuracy 1.00 78704730
macro avg 0.97 1.00 0.98 78704730
weighted avg 1.00 1.00 1.00 78704730
Confusion Matrix:
Predicted
O B-FUNC E-FUNC
Actual O 78095914 15481 13141
B-FUNC 1673 297051 272
E-FUNC 1979 153 279066
-----------------------------
--- Instruction Boundary Classification Report ---
Class Distribution:
NOT-START: 60,166,358 (76.45%)
INST-START: 18,538,372 (23.55%)
precision recall f1-score support
NOT-START 1.00 1.00 1.00 60166358
INST-START 1.00 1.00 1.00 18538372
accuracy 1.00 78704730
macro avg 1.00 1.00 1.00 78704730
weighted avg 1.00 1.00 1.00 78704730
Confusion Matrix:
Predicted
NOT-START INST-START
Actual NOT-START 60104092 62266
INST-START 13612 18524760
-----------------------------
