CONTEXTUAL VALIDATION OF RISK-RELATED FEATURES IN FORENSIC ANALYSIS SYSTEMS USING ARTIFICIAL INTELLIGENCE

Automated systems and language models in digital forensics select material from mobile-device exports for expert review. This paper adds contextual validation of an identified feature–artifact pair to this process. The method helps distinguish a person's account of their actions from a quotation. It examines the fragment alongside related messages and data, checks the person's role, the fragment's exact source location and provenance, and whether the feature warrants expert review. If the model cannot determine whether the material supports the feature, the result is recorded as undetermined. Otherwise, selection requires all checks to pass. Selected material includes a source reference. The experiment used 600 synthetic cases grouped into 300 pairs of comparable cases: similar fragments requiring different decisions because of context, attribution, or provenance. It covered 20 primary features and five context types. Against the expected decisions defined during dataset generation, full-context analysis incorrectly selected 155 of 300 cases lacking grounds for selection. Checks applied to the same responses reduced this count to 41. Among 300 supported cases, selection decreased from 297 to 289. Overall, 330 cases were selected, 241 rejected, and 29 recorded as undetermined. These results concern one local model and predefined metadata.