CyBERTuned-SecurityLLM

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

CyBERTuned

CyBERTuned is a BERT-like model trained with an NLE (non-linguistic element) aware pretraining method tuned for the cybersecurity domain.

Sample Usage

>>> from transformers import pipeline
>>> folder_dir = "CyBERTuned"
>>> unmasker = pipeline('fill-mask', model=folder_dir)
>>> unmasker("RagnarLocker, LockBit, and REvil are types of <mask>.")

[{'score': 0.8489783406257629, 'token': 25346, 'token_str': ' ransomware', 'sequence': 'RagnarLocker, LockBit, and REvil are types of ransomware.'},
{'score': 0.1364559829235077, 'token': 16886, 'token_str': ' malware', 'sequence': 'RagnarLocker, LockBit, and REvil are types of malware.'},
{'score': 0.0022238395176827908, 'token': 1912, 'token_str': ' attacks', 'sequence': 'RagnarLocker, LockBit, and REvil are types of attacks.'},
{'score': 0.001197474543005228, 'token': 11341, 'token_str': ' infections', 'sequence': 'RagnarLocker, LockBit, and REvil are types of infections.'},
{'score': 0.0009669850114732981, 'token': 6773, 'token_str': ' files', 'sequence': 'RagnarLocker, LockBit, and REvil are types of files.'}]

>>> # text requiring url comprehension (redirection attack), modified from https://intezer.com/blog/research/targeted-phishing-attack-against-ukrainian-government-expands-to-georgia/
>>> url_text = 'The PDF contains an action object. Upon a victim opening the PDF it will send a query to Google: http://www[.]google[.]com/url?q=http%3A%2F%2F9348243249382479234343284324023432748892349702394023.xyz&sa=D&sntz=1&usg=AFQjCNFWmVffgSGlrrv-2U9sSOJYzfUQqw. This link is a typical <mask> attack.'
>>> unmasker(url_text)[0]

{'score': 0.1701660305261612, 'token': 30970, 'token_str': ' redirect', 'sequence': 'The PDF contains an action object. Upon a victim opening the PDF it will send a query to Google: http://www[.]google[.]com/url?q=http%3A%2F%2F9348243249382479234343284324023432748892349702394023.xyz&sa=D&sntz=1&usg=AFQjCNFWmVffgSGlrrv-2U9sSOJYzfUQqw. This link is a typical redirect attack.'}

>>> from transformers import AutoModel, AutoTokenizer
>>> model = AutoModel.from_pretrained(folder_dir)
>>> tokenizer = AutoTokenizer.from_pretrained(folder_dir)
>>> text = "Cybersecurity information is often technically complex and relayed through unstructured text, making automation of cyber threat intelligence highly challenging."
>>> encoded = tokenizer(text, return_tensors="pt")
>>> output = model(**encoded)
>>> output[0].shape

torch.Size([1, 27, 768])

Citation

If you're using CyBERTuned please cite the following paper:

Eugene Jang, Jian Cui, Dayeon Yim, Youngjin Jin, Jin-Woo Chung, Seungwon Shin, and Yongjae Lee. 2024. Ignore Me But Don’t Replace Me: Utilizing Non-Linguistic Elements for Pretraining on the Cybersecurity Domain. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 29–42, Mexico City, Mexico. Association for Computational Linguistics.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0006
• train_batch_size: 64
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 2048
• total_eval_batch_size: 128
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-06
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.048
• num_epochs: 200

Framework versions

• Transformers 4.27.0.dev0
• Pytorch 1.12.1
• Datasets 2.6.1
• Tokenizers 0.13.2