The Attribution Problem
Every day, organizations around the world face cyberattacks ranging from ransomware campaigns to sophisticated espionage operations. When a breach occurs, one of the most critical and most difficult questions to answer is: who did this?”
Cyber threat attribution is the process of identifying the individual, group, or nation-state responsible for a cyberattack. It underpins everything from incident response decisions to geopolitical policy. Without the ability to attribute an attack, you cannot deter future ones, pursue legal action, or adapt your defenses to counter the specific adversary targeting you.
Yet traditional attribution remains extraordinarily difficult. Threat actors use proxy infrastructure, stolen credentials, false flags, and shared tooling to obscure their identity. The MITRE ATT&CK framework catalogues hundreds of techniques, and many are used by multiple groups. A single indicator of compromise (IOC), an IP address, a malware hash, or a domain name usually points to multiple actors.
This aspect is where artificial intelligence enters the picture.
The Limits of Manual Attribution
Traditional attribution workflows rely heavily on human analysts correlating threat intelligence from multiple sources. An analyst might examine malware samples, network traffic patterns, command-and-control infrastructure, phishing lure language, victimology, and geopolitical context to build an attribution hypothesis.
This approach has several fundamental limitations:
Scale. The volume of threat data generated daily far exceeds what human analysts can process. In 2025 alone, IBM X-Force reported a 49% increase in active ransomware groups, with smaller transient operators making attribution even more complex.
Speed. By the time an analyst completes a thorough attribution analysis, the threat actor may have already pivoted infrastructure, changed tactics, or launched additional campaigns.
Consistency. Different analysts examining the same evidence may reach different conclusions. Attribution confidence levels are inherently subjective, and cognitive biases, such as anchoring on familiar threat actors, for instance, can skew results.
Adversarial evasion. Sophisticated actors deliberately plant false flags. Russian APT groups have been observed embedding Chinese-language strings in malware. North Korean operators have mimicked Iranian TTPs. Manual analysis is particularly vulnerable to these deception techniques, as human analysts may overlook embedded strings or patterns that AI could easily detect.
How AI Transforms Attribution
AI and machine learning offer a fundamentally different approach to attribution, one that can process vast datasets, identify subtle patterns invisible to human analysts, and do so at machine speed.
Natural Language Processing for Threat Intelligence
A significant portion of threat intelligence exists as unstructured text: reports from vendors, dark web forum posts, malware analysis blogs, government advisories, and leaked communications. NLP models can automatically extract structured data from these sources, including threat actor names, malware families, targeted sectors, geographic indicators, and TTPs mapped to the MITRE ATT&CK framework.
More importantly, NLP can identify linguistic patterns that may indicate authorship. The coding style in malware, the language used in phishing lures, and even the comment conventions in exploit code can serve as behavioral fingerprints. Transformer-based models trained on large corpora of threat intelligence can learn to associate these linguistic features with known threat actor clusters.
Graph Neural Networks for Relationship Mapping
Cyberattacks don’t happen in isolation. Threat actors reuse infrastructure, share tools, operate within overlapping networks, and target similar victims. These relationships form a graph, and graph neural networks (GNNs) are uniquely suited to analyzing them.
A threat attribution graph might model:
- Nodes: IP addresses, domains, malware samples, email addresses, threat actors, campaigns, targeted organisations
- Edges: “communicates with,” “hosted on,” “authored by,” “targets,” “uses technique”
GNNs can learn node embeddings that capture both local features (what an IP address does) and structural features (how it relates to other entities in the graph). When a new attack is observed, the model can compute similarity between the attack’s subgraph and known threat actor subgraphs, generating attribution hypotheses ranked by confidence.
Recent studies have shown that GNNs can greatly change how we understand this area, allowing us to group attack campaigns, create ideas about who is behind them, and find new connections between threat actors.
Behavioural Pattern Analysis
Beyond static indicators, AI models can analyze behavioral patterns and the sequence of actions an attacker takes within a compromised network. This includes:
- Lateral movement patterns: How does the actor move through a network? Do they favor certain tools (PsExec, WMI, RDP)?
- Data exfiltration methods: Do they stage data before exfiltration? What protocols do they use?
- Temporal patterns: What hours do they operate? Are there consistent time zones?
- Tool preferences: Do they bring their tools or live off the land? Which LOLBins do they prefer?
Machine learning models trained on labeled campaign data can learn to cluster these behavioral sequences, identifying consistent patterns even when surface-level indicators change.
A Proposed Framework: The AI Attribution Pipeline
Based on my research at the University of Southampton under Dr. Erisa Karafili, I’m developing a framework for AI-driven threat attribution that combines multiple analytical approaches into a unified pipeline:
Stage 1: Data Ingestion and Normalisation
Ingest threat data from multiple sources, OSINT feeds, STIX/TAXII streams, internal SIEM logs, malware sandboxes, and dark web monitoring, and normalize it into a common format. This stage uses NLP to extract structured IOCs and TTPs from unstructured reports.
Stage 2: Feature Engineering
Extract features across multiple dimensions:
- Technical features: Malware binary attributes, network communication patterns, exploitation techniques
- Behavioural features: TTP sequences mapped to ATT&CK, dwell time, lateral movement paths
- Linguistic features: Code comments, phishing lure analysis, communication style
- Infrastructure features: Hosting providers, registration patterns, SSL certificate attributes, DNS behaviour
Stage 3: Graph Construction
Build a knowledge graph connecting all entities and their relationships. This graph is continuously updated as new intelligence arrives, creating a living model of the threat landscape.
Stage 4: Multi-Model Attribution
Apply multiple ML (machine learning) models in parallel:
- Supervised classifiers trained on labelled campaign data for known threat actors
- GNNs operating on the knowledge graph for structural similarity analysis
- Clustering algorithms for identifying previously unknown threat actor groups
- Anomaly detection for flagging attacks that don’t match any known actor profile
Stage 5: Confidence Scoring and Explanation
Each model produces an attribution hypothesis with a confidence score. The pipeline aggregates these into a final assessment, weighted by model reliability and evidence quality. Crucially, the system provides explainable outputs highlighting which specific features and relationships drove the attribution decision.
This explainability is essential. Attribution decisions can have real-world consequences, from sanctions to military responses. A black-box “it was Country X” isn’t actionable. Security teams need to understand why the model reached its conclusion, which evidence is strongest, and where uncertainty remains.
Challenges and Open Questions
AI-driven attribution is not a solved problem. Several significant challenges remain:
Data Quality and Ground Truth
ML models are only as accurate as their training data. Attribution ground truth is inherently uncertain; even the most confident human attributions involve some degree of speculation, which can lead to inaccuracies in the training data used for ML models. Training models on potentially incorrect labels can propagate and amplify errors, leading to significant misclassifications and undermining the reliability of the model’s predictions.
Adversarial Robustness
If threat actors know AI is being used for attribution, they can deliberately manipulate the features these models rely on. Adversarial machine learning crafting inputs designed to fool classifiers is a well-studied field, and attribution models must be hardened against such attacks.
Class Imbalance
Some threat actors are extensively documented (APT28, APT29, Lazarus Group), while others have minimal public reporting. Models may become biased toward well-documented actors, misattributing attacks by lesser-known groups, which can lead to significant consequences for the security landscape and misinformed responses from organizations.
Ethical and Legal Considerations
Automated attribution raises serious questions about accountability. If an AI system attributes an attack to a nation-state and that attribution informs a policy response, who bears responsibility if the attribution is wrong? The EU AI Act and similar regulatory frameworks will likely classify such systems as high-risk, requiring transparency and human oversight.
Cross-Domain Fusion
The most effective attribution requires fusing technical evidence with geopolitical context, human intelligence, and strategic analysis. Pure ML approaches struggle with this kind of contextual reasoning, suggesting that the optimal approach is human-AI collaboration rather than full automation, as this collaboration can leverage human insights to enhance the accuracy and relevance of the analysis in complex geopolitical situations.
The Road Ahead
The cybersecurity community is currently facing a pivotal moment. The volume and sophistication of cyberattacks are growing faster than human analyst capacity can scale. AI-driven attribution isn’t a luxury; it’s becoming a necessity.
In the near term, I expect to see:
- Graph-based approaches becoming mainstream in threat intelligence platforms, with tools like OpenCTI integrating GNN-powered analytics
- LLM-augmented analysis where large language models assist analysts in synthesising threat reports and generating attribution hypotheses
- Federated attribution models where organisations collaborate on model training without sharing sensitive incident data
- Standardised confidence frameworks that provide consistent, comparable attribution confidence scores across the industry
My ongoing research aims to contribute to this future by building practical, understandable, and robust AI systems that can help defenders answer the critical question of “who” with greater speed, accuracy, and confidence than ever before.
This post is part of my ongoing PhD research Proposal on AI-driven cybersecurity threat attribution at the University of Southampton. I’ll be publishing follow-up posts diving deeper into specific techniques, including a hands-on walkthrough of building a threat attribution knowledge graph.
If you’re working in threat intelligence or attribution research, I’d love to connect reach out via the contact page or find me on LinkedIn.