StriderSPD

Repository for the TSE paper "StriderSPD: Structure-Guided Joint Representation Learning for Binary Security Patch Detection".

Install

Use Linux, Python 3.11, and CUDA GPUs supporting BF16. The original execution uses four visible GPUs in one Python process; the code uses model parallelism.

git clone https://github.com/7thParkk/StriderSPD.git
cd StriderSPD
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

The environment pins PyTorch 2.4.0 with CUDA 12.1, Transformers 4.55.0, Accelerate 1.10.0 and PyG 2.6.1. Allow disk space for the checkpoint and base-model files, and RAM for loading the full checkpoint on CPU.

Download the trained model

The paper joint checkpoint is hosted at hanjin996/StriderSPD:

• model_epoch5_best_acc_qwen3_8b_824.pth, 25,682,301,190 bytes (25.68 GB).
• benchmark_1720.zip, an alternative download of the benchmark already included here.

Download and verify the checkpoint:

python scripts/download_checkpoint.py --output-dir checkpoints

SHA256: 991045a6bd0c8a30b113b1d2f633ed6d464f48ec06f4325b6adfef27db5e01e7.

The checkpoint stores the complete joint model, including the Qwen3-8B backbone, UniXcoder node encoder, GNN, attention and classifier. Evaluation uses the public Qwen/Qwen3-8B and microsoft/unixcoder-base assets to initialize the custom architecture and tokenizers, then strictly replaces all model tensors with the checkpoint. Local model directories can be supplied with --pretrained_llm_path and --unixcoder_path.

Benchmark

benchmark_1720/
├── benchmark_new_infos.json
├── CFG/<project>/<optimization>/<file>.json
└── pseudo/<project>/<optimization>/<file>.json

There are 1,720 paired samples: 1,010 security and 710 non-security. Each modality has 3,440 pre/post JSON files. Projects, optimization levels, file names, manifest order and JSON contents are preserved from the original benchmark. CFG JSON contains nodes and edges; pseudo JSON contains a pseudo_code list. Labels and pair paths are in the manifest.

Hard-negative sample IDs (the id field in benchmark_new_infos.json) are listed in hard_negative_ids.txt.

Check all data representations without ML dependencies:

python scripts/check_benchmark.py

When using the ZIP download, extract it into benchmark_1720/:

unzip benchmark_1720.zip -d benchmark_1720

Inference / evaluation

Run from the repository root:

CUDA_VISIBLE_DEVICES=0,1,2,3 python src/main.py \
  --do_test \
  --checkpoint_path checkpoints/model_epoch5_best_acc_qwen3_8b_824.pth \
  --llm_architecture_type qwen3 \
  --attention_layer_frequency 1 \
  --root_path benchmark_1720 \
  --ass_path benchmark_1720/CFG \
  --pse_path benchmark_1720/pseudo \
  --eval_batch_size 1 \
  --output_dir outputs/eval

Add --preflight to validate the manifest, all required files and checkpoint path before loading a model. Evaluation requires all 1,720 samples and fails when input files are missing. Results are saved to outputs/eval/predictions.jsonl in manifest order and outputs/eval/summary.json. Positive means a security patch; prediction uses sigmoid(logit) >= 0.5.

The three GNN roles default to cuda:1,cuda:2,cuda:3. To run with one sufficiently large GPU, use CUDA_VISIBLE_DEVICES=0 and --gnn_devices cuda:0,cuda:0,cuda:0; memory sufficiency and GPU result parity for this layout have not been measured here. Use python src/main.py --help for the complete CLI.

For another evaluation set, provide the same manifest/representation schema, use --eval_split <name> to load <name>_infos.json, and set --expected_eval_samples <count>.

Joint training

The joint-training entrypoint expects a prepared training dataset and a Qwen3-8B initializer. The paper run initialized from a Qwen3-8B model with its semantic fine-tuning LoRA weights already merged. Supply that local merged model to reproduce the joint-training configuration.

training_data/
├── train_infos.json
├── dev_infos.json
├── CFG/<project>/<optimization>/<file>.json
└── pseudo/<project>/<optimization>/<file>.json

CUDA_VISIBLE_DEVICES=0,1,2,3 python src/main.py \
  --do_train \
  --pretrained_llm_path /path/to/qwen3-8b-lora-merged \
  --llm_architecture_type qwen3 \
  --attention_layer_frequency 1 \
  --root_path /path/to/training_data \
  --ass_path /path/to/training_data/CFG \
  --pse_path /path/to/training_data/pseudo \
  --batch_size 2 \
  --eval_batch_size 2 \
  --gradient_accumulation_steps 1 \
  --epoch 8 \
  --learning_rate 1e-5 \
  --weight_decay 1e-4 \
  --adam_epsilon 1e-6 \
  --max_grad_norm 1.0 \
  --seed 44 \
  --save_every_epoch \
  --save_path outputs/checkpoints

The 1,720-sample release supplies evaluation data. The train/dev representations and semantic-stage merged initializer must be prepared separately. Training selects the best validation-accuracy checkpoint and records its path in best_checkpoint.txt; --save_every_epoch also retains each epoch checkpoint.

Paper reference

Paper checkpoint reference results on the 1,720-sample benchmark:

Samples	Accuracy	Precision	Recall	F1	FPR
1,720	0.854070	0.822979	0.957426	0.885126	0.292958