AI-Based Smart Communication for Digital Forensics and Crime Analysis
Year: 2026
Type: conference-paper
Abstract: Digital forensics increasingly utilizes local language models for automated triage, yet these systems often evaluate text strings in isolation from their physical storage location. This context-blindness leads to high false-positive rates and reported error rate, as benign system code is frequently misidentified as evidence. This research proposes the Geo-Contextualized ForensicLLM framework to address the limitations of file system metagraph blindness. The methodology implements a six-step ingestion pipeline that utilizes a dual stream parser and a spatial math encoder to serialize directory depth and zone boundaries. These spatial anchors are injected as prefix tokens, forcing attention-constrained inference within the model. Experimental validation using a 10,000 file benchmark dataset demonstrates a precision of 91.98% and an F1-Score of 94.14%. The system achieved an observed 100% Attribution Grounding Rate (AGR), measurably reducing false positive backlogs compared to context-blind models such as ForensicLLM. This approach ensures verifiable evidence anchoring without requiring additional neural parameters, supporting evidentiary integrity in air-gapped environments. The framework demonstrates resilience against specific adversarial counter-forensic tactics and evidentiary hallucination risks.
Language: English
Topic: Digital and Cyber Forensics
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203449727