DiffuPac: Contextual Mimicry in Adversarial Packets Generation via Diffusion Model
Authors: Abdullah Bin Jasni, Akiko Manada, Kohei Watabe
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments to validate the performance of Diffu Pac on 6 types of attacks, against 6 classifiers. |
| Researcher Affiliation | Academia | Abdullah Bin Jasni Graduate School of Engineering Nagaoka University of Technology Nagaoka, Japan s203108@stn.nagaokaut.ac.jp Akiko Manada Graduate School of Engineering Nagaoka University of Technology Nagaoka, Japan amanada@vos.nagaokaut.ac.jp Kohei Watabe Graduate School of Science and Engineering Saitama University Saitama, Japan kwatabe@mail.saitama-u.ac.jp |
| Pseudocode | Yes | Algorithm 1 Finding contextually relevant packet sequences. |
| Open Source Code | No | Due to ethical concerns, we have decided not to publicly share the data and code. Additionally, much of the code is based on existing codebases from other sources. |
| Open Datasets | Yes | The datasets used for this model are Kitsune Dataset (Mirsky et al. [2018]) and CICIDS2017 Dataset (Sharafaldin et al. [2018]). |
| Dataset Splits | Yes | Initially, pre-training of the BERT model utilizes a large subset of unlabeled network traffic to leverage the model s capability to capture diverse traffic patterns, accounting for 60% of the total data. Fine-tuning phase, the focus shifts to a smaller, labeled dataset, which constitutes 20% of the total data. This dataset is distinctly partitioned into malicious and normal packets, which is crucial for training the model to mimic malicious packets as normal. Training the classifier and NIDS then utilizes another 10% of the total data consisting of labeled portion. Testing phase is conducted with the remaining 10% of the data, reserved exclusively for evaluating the model s efficacy. |
| Hardware Specification | Yes | All training, experimentation, and sampling processes are executed on a single NVIDIA AD102 (Ge Force RTX 4090) GPU. |
| Software Dependencies | No | The paper mentions software like FAISS library, UER-py project, and Diffu Seq codebase but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | Table 4: Parameters used for pre-training the BERT model. Table 6: Parameters used for the diffusion model. |