asmdetect

Compiler-Aware Assembly-Level Malware Detection using Fine-Tuned Language Models

A Python library for static malware detection using x86 assembly opcode sequences. Fine-tunes microsoft/codebert-base on disassembled binaries to classify files as malware or benign — no execution required.

Built for SOC teams who need fast, explainable triage of suspicious binaries before deeper sandbox analysis.

---

Model

HuggingFace: Swarnadharshini/codebert-malware-detector

Property	Value
Base model	microsoft/codebert-base
Task	Binary sequence classification
Input	x86 opcode sequences (max 512 tokens)
Output	malware / benign + confidence score
Parameters	~125M (8 frozen layers + 4 fine-tuned + classifier)
Training data	1,458 samples (with augmentation)
Dataset	Arun152k — objdump-disassembled PE binaries
Decision threshold	0.62 (calibrated on test set)

Dataset

Processed opcode sequences available on Kaggle: https://www.kaggle.com/datasets/swarnadharshini/malware-opcodes

Original source: Arun152k/Malware-Detection-using-N-Gram-Frequency

Test set results

Metric	Value
Accuracy	86.0%
F1 Score	0.857
Precision	86.8%
Recall	84.6%
AUC-ROC	0.910
False Negatives	12
False Positives	10

---

Installation

pip install -r requirements.txt
pip install -e .

---

Quick start

from asmdetect import MalwareDetector

# Load model from HuggingFace Hub (downloads once, cached locally)
detector = MalwareDetector.from_pretrained()

# Predict from a .csv assembly file (IDA Pro / objdump format)
result = detector.predict_file("suspicious_binary.csv")
print(result)
# DetectionResult(
#   source      = suspicious_binary.csv
#   prediction  = MALWARE  [HIGH] HIGH
#   confidence  = 99.60%
#   malware_p   = 99.60%
#   benign_p    = 0.40%
#   tokens      = 312  (truncated=False)
#   threshold   = 0.62
# )

# Predict from a raw opcode string
result = detector.predict_text("push mov xor xor call ret add nop")

# Batch triage a folder — sorted by malware probability (highest first)
results = detector.predict_batch("/soc/incoming/suspicious/")
for r in results[:5]:
    print(f"{r.malware_probability*100:.1f}%  {r.risk_level}  {r.source}")

---

CLI

# Single file
asmdetect --file suspicious.csv

# Raw opcode string
asmdetect --text "push mov xor call ret"

# Batch triage a folder
asmdetect --batch /soc/incoming/

# Custom threshold
asmdetect --file suspicious.csv --threshold 0.70

# JSON output (for SIEM/SOAR integration)
asmdetect --file suspicious.csv --json

---

Scripts

# Run demo predictions from HuggingFace
python scripts/predict_from_hf.py

# Predict a specific file
python scripts/predict_from_hf.py --file path/to/sample.csv

# Predict raw opcodes
python scripts/predict_from_hf.py --text "push mov xor call ret"

# SOC batch triage with JSON report
python scripts/batch_triage.py --folder /soc/incoming/ --report report.json

---

API reference

MalwareDetector

MalwareDetector.from_pretrained(
    model_id  = "Swarnadharshini/codebert-malware-detector",  # or local path
    threshold = 0.62,   # decision threshold — lower = more sensitive
    device    = "auto", # 'cuda', 'cpu', or 'auto'
)

Method	Description
predict_file(filepath)	Classify a .csv assembly file
predict_text(text)	Classify a raw opcode string
predict_batch(folder)	Classify all CSVs in a folder, sorted by risk
set_threshold(value)	Update threshold at runtime
benchmark(text, n=10)	Measure inference latency

DetectionResult

Field	Type	Description
prediction	str	'malware' or 'benign'
label	int	1 = malware, 0 = benign
confidence	float	Probability of predicted class
malware_probability	float	Raw P(malware)
benign_probability	float	Raw P(benign)
risk_level	str	'HIGH' ≥80% | 'MEDIUM' ≥55% | 'LOW'
input_tokens	int	Token count before truncation
truncated	bool	True if input exceeded 512 tokens
is_malware	bool	Convenience property
is_benign	bool	Convenience property
to_dict()	dict	Serialise to dictionary
to_json()	str	Serialise to JSON string

Threshold tuning

# High-security: catch more malware (more false alarms)
detector.set_threshold(0.45)

# Low-noise: fewer false alarms (may miss borderline cases)
detector.set_threshold(0.75)

# Reset to calibrated default
detector.set_threshold(0.62)

---

Input format

CSV files — from IDA Pro or objdump -d:

Address,Hex_Opcode,Opcode,Operand 1,Operand 2
402000:,30 7d 07,xor,%bh,
402003:,00 00,add,%al,

Raw opcode strings:

detector.predict_text("push mov sub lea call add pop ret")

---

Project structure

asmdetect/
├── asmdetect/
│   ├── __init__.py        ← public API
│   ├── detector.py        ← MalwareDetector class
│   ├── preprocessing.py   ← opcode extraction utilities
│   ├── result.py          ← DetectionResult dataclass
│   ├── cli.py             ← asmdetect CLI
│   └── version.py
├── tests/
│   └── test_asmdetect.py  ← 26 unit tests
├── scripts/
│   ├── predict_from_hf.py ← standalone HF demo script
│   └── batch_triage.py    ← SOC batch triage with report
├── examples/
│   └── basic_usage.py     ← usage examples
├── notebooks/
│   └── malware-detect.ipynb ← Kaggle training notebook
├── pyproject.toml
├── requirements.txt
└── README.md

---

Training

The model was trained on Kaggle (Tesla P100 GPU) using the pipeline in notebooks/malware-detect.ipynb.

Training iterations:

Run	Key changes	Accuracy	F1	AUC-ROC
Baseline	Default cross-entropy loss	78.9%	0.802	0.890
Run 2	Frozen layers + weighted loss (2.0)	76.4%	0.800	0.854
Run 3	weight=1.5 + threshold calibration + augmentation	86.0%	0.857	0.910

---

Limitations

• Trained on 1,042 samples — a larger dataset will improve generalisation
• Input truncated to 256 opcodes; signatures in binary tails may be missed
• Obfuscated/packed malware that alters opcode distribution may evade detection
• Designed for x86 Windows PE binaries; accuracy on ARM or ELF is untested
• Static analysis only — does not detect runtime/memory-resident malware

AsmDetect

Assembly-level malware detection using CodeBERT.

Install

pip install asmdetect

Usage

from asmdetect import MalwareDetector

detector = MalwareDetector.from_pretrained()
detector.predict_text("push mov xor call")