TL;DR
This paper explores using Large Language Models (LLMs) like GPT-4 and embedding models to automatically identify threat actor Tactics, Techniques, and Procedures (TTPs) from cyber threat intelligence (CTI) reports. They developed an end-to-end pipeline that can predict threat actors based on these extracted TTPs, showing promising results that outperform random guessing, even with noisy LLM outputs.
What Problem Does This Paper Address?
Attributing cyberattacks to specific threat actors is a crucial but incredibly challenging task for cybersecurity defenders. The process typically involves manually sifting through vast amounts of forensic documentation to identify an adversary's TTPs. This manual effort is time-consuming, leading to significant delays in attribution, which can have serious geopolitical consequences. Existing automated methods often rely on low-level indicators that are easily spoofed, or require extensive manual data preparation. This research aims to automate the TTP extraction and attribution process using modern NLP techniques, specifically LLMs, to speed up and improve the accuracy of cyber-attack attribution.
Methods
The researchers developed a two-step pipeline:
- TTP Identification: This step focuses on extracting TTPs from raw CTI documents. They explored two main approaches:
- LLM Prompting: Using a pre-trained LLM (GPT-4) with a carefully crafted prompt to directly extract TTPs from a document.
- Vector Embedding Search: This method involves creating vector embeddings for chunks of text within a document and comparing them to pre-computed embeddings of known MITRE ATT&CK TTPs. The TTP with the highest cosine similarity is selected. They also experimented with the "Hypothetical Document Embeddings" (HyDE) approach to augment TTP definitions.
- Threat Actor Attribution: Once TTPs are identified (either by LLM or vector embedding), this step uses them to predict the threat actor. They trained a machine learning model using these TTPs. The model learns a "weights matrix" representing the likelihood of specific TTPs being associated with particular threat actors. For a new document, TTPs are extracted, and this matrix is used to generate a ranked list of potential threat actors.
Datasets
- MITRE ATT&CK Groups Page: This was used as a proxy for human-generated TTP datasets. The researchers leveraged the references (raw post-incident reports) and associated TTPs listed on MITRE's group pages.
- Open-sourced Reports: Additional reports were gathered for 12 of the 29 threat actors to augment the MITRE data.
- Training/Validation/Test Splits: The MITRE ATT&CK dataset was split into 70% for training, 20% for validation, and 10% for testing to develop and evaluate the attribution model.
Key Insights and Ideas
- LLMs as a "Noisy" but Useful Tool: While LLMs like GPT-4 generate TTP datasets with significant "noise" (i.e., TTPs not directly found in human-annotated datasets), these outputs still retain frequency patterns that align with established TTPs. This suggests that LLMs can be valuable for initial TTP extraction, even if human review or further processing is needed.
- Vector Embeddings Offer Transparency: The vector embedding approach, while sometimes producing more specific TTPs, offers greater transparency than the "black-box" nature of LLMs. The probabilistic weights generated by this method provide a more nuanced view of TTP associations compared to binary inclusion/exclusion.
- "Attributability" Varies by Threat Actor: The study found that some threat actors are more "attributable" than others. Groups with highly unique TTP profiles and less overlap with other actors' TTPs are easier to identify. Conversely, actors with very broad or similar TTP sets are harder to distinguish.
- Expert Priors Improve Performance: Incorporating an "expert prior" (e.g., analyst's knowledge about the prevalence of certain threat actors) significantly improves the attribution model's performance, leading to better average rankings.
Key Results
- Jaccard Similarity: LLM-generated TTPs showed an average Jaccard similarity of 39% ± 12% with human-generated MITRE data, while vector embedding search yielded 18% ± 7.39%. Both are considered low, but the LLM approach performed better in terms of direct overlap.
- Set Differences: GPT-4 omitted about 41% of MITRE TTPs ("In MITRE, not GPT") and generated about 66% of TTPs not found in MITRE ("In GPT, not MITRE"), indicating noise. The vector embedding approach showed similar noise levels.
- Attribution Performance: The best-performing attribution model achieved an average rank of 7.55 out of 29 threat actors, significantly outperforming random guessing (average rank of 15). This demonstrates the effectiveness of TTP-based attribution even with imperfect TTP extraction.
- Hallucination Rate: GPT-4's hallucination rate (TTPs not mapping to MITRE) was low, at less than 1% overall.
Limitations and Future Work
- LLM Noise: The primary limitation acknowledged is the "noise" in LLM-generated TTPs, which can make direct comparison difficult and requires careful handling.
- Data Volume: The study suggests that providing more data for TTP generation leads to better attribution, implying that more comprehensive CTI reports would yield improved results.
- Actor Similarity: Threat actors with highly similar TTPs or broad TTP sets are harder to attribute, highlighting a limitation of technique-based attribution for certain groups.
- Future Work: The authors suggest exploring fine-tuning LLMs with human feedback to improve TTP extraction, incorporating dynamic threat actor relevance signals, and further evaluating the pairwise similarity between LLM and vector embedding TTPs. They also emphasize the need for further research into how these automated tools can best augment human analysts.
Summary updated