SROBR: Semantic Representation of Obfuscation-Resilient Binary Code

Table 7

Experimental results of model variants.

options	sub	bcf	fla
accuracy
Variants
Linear+GAT	0.785	0.675	0.651
LSTM+GAT	0.759	0.577	0.531
BiLSTM+GAT	0.838	0.676	0.682
BERT+GCN	0.852	0.608	0.648
SROBR	0.888	0.701	0.694

Linear+GAT removes the BERT module, and the instruction vector is embedded only by random initialization. It is used to explore the role of the BERT module.
LSTM+GAT uses LSTM instead of BERT to generate the instruction embeddings in the basic block, which is used to compare the effects of BERT and RNN.
BiLSTM+GAT uses BiLSTM to further explore the pros and cons of recurrent neural networks and BERT.
BERT+GCN replaces the GAT module with GCN without using attention weights.