Forensic Evidence Prompt Injection: Anti-Forensic Attacks Against LLM-Assisted Digital Forensics
Start here
Researchers propose a taxonomy and proof of concept for anti-forensic attacks against LLM-assisted digital forensics: the analyzed evidence is controlled by the attacker and must be read closely by forensic personnel, making the model's reasoning layer a new anti-forensic target; attacks can be combined by objective, technique, payload location, and concealment method, and are invisible to forensic personnel. The authors demonstrate three attacks on an LLM-assisted analysis workflow; planting text only on a seized device can change a report's conclusions and write an attack to an analyst's workstation.
Generated by AI from reports · Updated 16 days ago
Related research
This event's paper
Progress timeline
- Date not givenarXivPrompt injection through forensic evidence: a taxonomy and proof of concept of anti-forensic attacks against LLM-assisted digital forensics
Researchers propose a taxonomy and proof of concept of anti-forensic attacks against LLM-assisted digital forensics. Digital forensics inverts the usual premise of indirect prompt injection: the evidence being analyzed is by definition controlled by the attacker, and forensic personnel must read it carefully, making the model reasoning layer a new anti-forensic target beyond data and tools. Attackers can construct attacks by combining target, technique, payload location and concealment method; perceptual differences between deterministic tools, forensic personnel and the model regarding the same evidence make the injection invisible to forensic personnel. The authors demonstrate three attacks on an LLM-assisted analysis workflow, where merely planting text on a seized device can change report conclusions, write attacker-specified content to the analyst workstation and trigger outbound network traffic, and they also propose mitigations.