TL;DR
The paper introduces the Forensic AI Life Cycle (FALC), a new framework that maps out the digital "breadcrumbs" left behind by AI systems. By shifting the focus from how AI is built to what artifacts it creates, FALC helps investigators find evidence and highlights the massive gaps where current forensic tools are currently "blind."
What Problem Does This Paper Address?
AI is increasingly appearing in criminal investigations as a tool (generating deepfakes), a target (model theft), or a source of evidence (autonomous vehicle logs). However, digital forensics hasn't kept pace.
Currently, investigators often don't know what traces to look for in a complex AI system. Furthermore, even when they find data, they often lack the established techniques to analyze it. Existing AI life cycle models are designed for process optimization or ethics, not for finding evidence after a crime. This paper aims to provide a systematic structure to locate evidence and a research agenda to fill the technical gaps in AI forensics.
Methods
The authors developed the FALC model through a conceptual synthesis of 20 existing AI life cycle models. Unlike traditional models that follow a linear development path, FALC is artifact-centric. It groups development steps based on the types of evidence they produce rather than when they occur.
The researchers then:
- Cataloged Artifacts: Identified specific traces (e.g., training checkpoints, prompt templates, audit results) for each stage.
- Mapped Crimes: Collected examples of AI-related crimes and the specific technical questions investigators need to answer.
- Created a Taxonomy: Classified these forensic tasks by whether they can be handled with existing tools, need adaptations, or require entirely new forensic science.
Datasets
While this is not a traditional machine learning paper, the authors used several "data" sources for their synthesis and validation:
- Life Cycle Models: A meta-analysis of 20 prominent AI life cycle models (e.g., CRISP-DM, TDSP).
- Case Study Validation: To test FALC's completeness, they applied it to the NTSB Highway Investigation HWY18MH010 (the 2018 Uber autonomous vehicle crash). They successfully mapped 30 different artifacts from the public docket to the FALC stages.
Key Insights and Ideas
- The "Donut" of Forensic Difficulty: Forensic difficulty is not evenly distributed. The "periphery" of the AI system (Governance, Deployment, and Data Acquisition) is largely tractable using established IT forensic methods. The "center"—the model itself and the massive datasets used to train it—is where current techniques fail.
- Governance as the Best Entry Point: Because AI models are often "black boxes," the authors suggest that investigators should start with the Governance and Quality Assurance stage. Artifacts here (audit results, version histories, design docs) are conventional documents that can corroborate findings from more complex stages.
- Non-Linearity is Key: In AI development, stages are often revisited or skipped. FALC reflects this by not imposing a strict order, encouraging investigators to look for artifacts from every stage regardless of where the suspected crime occurred.
- The Intent Gap: A major challenge identified is the inability to trace a model's behavior back to a specific human intention. It is currently nearly impossible to prove if a model's bias or "malicious" output was a deliberate design choice or an accidental emergence from the training data.
Key Results
The paper's primary output is a Taxonomy of AI Forensic Expertise (Table 1 in the paper). It identifies critical areas where "New Analysis Techniques" are urgently needed:
- Deepfake Detection: Still an active research problem with no "silver bullet" for investigators.
- Explainable AI (XAI) for Forensics: We need ways to prove why a model reached a conclusion in a way that is legally defensible in court.
- Linking Actions to User Input: Developing methods to prove that a specific user prompt directly caused a specific (and perhaps autonomous) AI action.
- Model Theft: Determining if a proprietary model was stolen or if a similar model was simply trained independently.
Limitations and Future Work
The authors acknowledge that FALC was derived conceptually and validated against only one major case study. While the model is designed to be extensible, the artifact catalog will need constant updates as AI technology evolves (e.g., new types of "agentic" AI).
The paper also explicitly excludes decommissioning (end-of-life) artifacts, such as deletion logs and certificates of destruction. The authors view "anti-forensics"—the deliberate destruction of evidence—as a separate, vital field for future research. Finally, the study relies primarily on English-language literature and industry practices, which may not capture global variations in AI development.
Summary updated