instruction stringclasses 1
value | input stringlengths 44 449 | output stringlengths 172 671 |
|---|---|---|
ForensicLLM: A local large language model for digital forensics | Why is it important that third parties (lukas et al., 2023)? | third parties (Lukas et al., 2023). General purpose cloud-based models like ChatGPT require internet connectivity, rendering them impractical Digital Forensics (DF) involves the acquisition, authentication, and in secure environments where internet access is restricted *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that also, while analysis of digital evidence in a manner that is both legal and scientific? | Also, while analysis of digital evidence in a manner that is both legal and scientific. these LLMs might offer broad applicability across various tasks, they A crucial aspect for investigators is ensuring the admissibility of digital often lack domain-specific expertise (Scanlon et al., 2023) *(Sharma et al.... |
ForensicLLM: A local large language model for digital forensics | What key issue is addressed in this section? | The evidence in court. The Daubert standards, established by the 1993 computing power requirements for training and inference can be sub- Daubert v *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that merrell dow pharmaceuticals case, set guidelines for evalu- stantial, leading to increased operational costs? | Merrell Dow Pharmaceuticals case, set guidelines for evalu- stantial, leading to increased operational costs. Fine-tuning these models ating the scientific validity and reliability of methodologies used to to specific tasks or domains can be challenging and time-consuming, gather evidence *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that the major daubert criteria include: 1) the ability to test requiring significant expertise and computational resources (sevilla the methodology and its prior testing, 2) known error rates, 3) peer et al., 2022; sharir et al., 2020). review and publication status, and 4) acceptance by the rel... | The major Daubert criteria include: 1) the ability to test requiring significant expertise and computational resources (Sevilla the methodology and its prior testing, 2) known error rates, 3) peer et al., 2022; Sharir et al., 2020). review and publication status, and 4) acceptance by the relevant sci- Smaller open-s... |
ForensicLLM: A local large language model for digital forensics | What key issue is addressed in this section? | These models accepted. can achieve comparable results to larger models when fine-tuned on Large Language Models (LLMs) often lack transparency regarding domain-specific datasets (Bolton et al., 2024; Rebei, 2023) *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that addition- their training data and internal mechanisms, complicating the verifica- ally, techniques such as quantization (jacob et al., 2018) and low-rank tion of their outputs’ scientific basis? | Addition- their training data and internal mechanisms, complicating the verifica- ally, techniques such as quantization (Jacob et al., 2018) and Low-Rank tion of their outputs’ scientific basis. Further, data confidentiality is Adaptation (LORA) (Hu et al., 2021) significantly reduce the computa- paramount in DF, and... |
ForensicLLM: A local large language model for digital forensics | Why is it important that concerns due to the potential exposure of sensitive case information to furthermore, rag approach (lewis et al., 2020) allows models to * corresponding author? | concerns due to the potential exposure of sensitive case information to Furthermore, RAG approach (Lewis et al., 2020) allows models to * Corresponding author. Baggil(i) Truth (BiT) Lab, Center for Computation & Technology, Baton Rouge, LA, USA. E-mail addresses: bsharm6@lsu.edu (B *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that published by elsevier ltd on behalf of dfrws? | Published by Elsevier Ltd on behalf of DFRWS. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). B *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | What key issue is addressed in this section? | Sharma et al. F o r e n s i c S c i e n c e I n t e r n a t i o n a l: D i g i t a l I n v e s t i g a t io n52(2025)301872 retrieve additiona... |
ForensicLLM: A local large language model for digital forensics | Why is it important that (2024)show that research in cybersecurity this section explores key background concepts integral to llms. and df has primarily focused on evaluating larger cloud-based models with minimal exploration of smaller, specialized models? | (2024)show that research in cybersecurity This section explores key background concepts integral to LLMs. and DF has primarily focused on evaluating larger cloud-based models with minimal exploration of smaller, specialized models. Our objective 3.1 *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that tokenization is to fill this research gap by examining the application of a domain-specialized llm in df? | Tokenization is to fill this research gap by examining the application of a domain-specialized LLM in DF. Specifically, our research questions are: Tokenization splits input text into smaller units called tokens, which can be words, subwords, or characters *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that llama-3 uses subword tokeni- rq1 how do fine-tuning and rag framework compare in terms of zation with the tiktoken byte pair encoding (bpe) tokenizer (ai@meta, performance in digital forensics? 2024a), containing over 128k unique tokens? | LLaMA-3 uses subword tokeni- RQ1 How do fine-tuning and RAG framework compare in terms of zation with the tiktoken Byte Pair Encoding (BPE) tokenizer (AI@Meta, performance in digital forensics? 2024a), containing over 128k unique tokens. This approach splits rare RQ2 How reliable is a fine-tuned LLM in retrieving re... |
ForensicLLM: A local large language model for digital forensics | Why is it important that token embeddings investigation? token embeddings are numerical vectors that capture the semantic meaning of the token and the relationships between tokens? | Token embeddings investigation? Token embeddings are numerical vectors that capture the semantic Our work makes the following key contributions: meaning of the token and the relationships between tokens. During inference, the token embedding layer in the LLM’s architecture acts as a • We introduce ForensicLLM, a mo... |
ForensicLLM: A local large language model for digital forensics | Why is it important that in most models, the token embedding layer is learned digital forensic artifacts? | In most models, the token embedding layer is learned digital forensic artifacts. To the best of our knowledge, this is the pri- through pretraining *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that in other cases, a pre-trained token embedding mary account for developing a local llm for df? | In other cases, a pre-trained token embedding mary account for developing a local LLM for DF. model is used such as word2vec (Mikolov et al., 2013) or Global Vectors • We evaluate ForensicLLM based on its performance on held-out test for Word Representation (GLOVE) (Pennington et al., 2014). datasets and its ability... |
ForensicLLM: A local large language model for digital forensics | Why is it important that attention 3.1–8b and llama-3.1–8b equipped with a rag component. • we conduct a user study with df professionals, evaluating model attention mechanisms are a key element of transformer-based llms responses for usefulness, correctness, relevance, citation, and such as llama and generati... | Attention 3.1–8B and LLaMA-3.1–8B equipped with a RAG component. • We conduct a user study with DF professionals, evaluating model Attention mechanisms are a key element of transformer-based LLMs responses for usefulness, correctness, relevance, citation, and such as LLAMA and Generative Pre-trained Transforme... |
ForensicLLM: A local large language model for digital forensics | Why is it important that enable the model to focus on relevant parts of the input sequence when generating outputs? | enable the model to focus on relevant parts of the input sequence when generating outputs. Specifically, the self-attention layer in transformers The paper is structured as follows: Section 2 covers related work, computes the relevance of each token in the input sequence to every Section 3 explores key LLM concepts... |
ForensicLLM: A local large language model for digital forensics | Why is it important that quantization optimizes llms by reducing the precision of numerical data, typically converting 32-bit floating-point weights to lower preci- scanlon et al? | Quantization optimizes LLMs by reducing the precision of numerical data, typically converting 32-bit floating-point weights to lower preci- Scanlon et al. (2023)evaluated ChatGPT for digital forensic tasks like sion formats (e.g., 16-bit or 8-bit) *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that this decreases memory usage and artifact analysis, evidence searching, and incident response? | This decreases memory usage and artifact analysis, evidence searching, and incident response. The accu- computational costs, enabling deployment on devices with limited re- racy, relevance, and completeness of the responses were analyzed *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that while sources without significantly compromising model accuracy (jacob chatgpt showed potential in generating regular expressions, enhancing et al., 2018). keyword searches, summarizing documents, and aiding in coding tasks, significant limitations were noted, such as biases from training d... | While sources without significantly compromising model accuracy (Jacob ChatGPT showed potential in generating regular expressions, enhancing et al., 2018). keyword searches, summarizing documents, and aiding in coding tasks, significant limitations were noted, such as biases from training data, lack 3.5. Sentence ... |
ForensicLLM: A local large language model for digital forensics | Why is it important that similar to token embeddings discussed in section 3.2, llms have also michelet and breitinger (2023) investigated the potential of using been designed to generate sentence embeddings, which capture the se- llms like chatgpt-3.5 and llama-2-13b to assist in digital forensic mantic meaning of l... | Similar to token embeddings discussed in Section 3.2, LLMs have also Michelet and Breitinger (2023) investigated the potential of using been designed to generate sentence embeddings, which capture the se- LLMs like ChatGPT-3.5 and LLaMA-2-13B to assist in digital forensic mantic meaning of longer text. These embeddi... |
ForensicLLM: A local large language model for digital forensics | Why is it important that chatgpt outperformed the locally run llama-2-13b large contexts, often composed of hundreds of tokens, with the primary model in generating more accurate and complete texts? | ChatGPT outperformed the locally run LLaMA-2-13B large contexts, often composed of hundreds of tokens, with the primary model in generating more accurate and complete texts. However, both goal of enhancing semantic information retrieval *(Sharma et al., 2025)* |
ForensicLLM: A local large language model for digital forensics | Why is it important that vector embedding models required significant human proofreading? | Vector embedding models required significant human proofreading. Further, they found models are trained so that semantically similar information will have that while LLMs cannot fully automate report writing, they can still help closer vector repres |
README.md exists but content is empty.
- Downloads last month
- 3