Leveraging Multimodal Large Language Models for Digital Forensics in Military Personnel Detection in Mobile Device Images

Taras Fedynyshyn, Olha Partyka, Ivan Opirskyy, Abilkair Imanberdi, Alisher Batkuldin, Olzhas Konakbayev

2025

Closed access Metadata (CSL-JSON)
DOI 10.1109/snpd65828.2025.11253944
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Digital forensics plays a crucial role in modern investigations, particularly in analyzing vast amounts of image data from mobile devices to identify individuals, objects, and activities of interest. Traditional forensic techniques struggle with the complexity of real-world images, especially in distinguishing between military personnel and mannequins dressed in military attire. Recent advancements in Artificial Intelligence (AI), particularly multimodal Large Language Models (LLMs), offer new opportunities for automating and enhancing forensic image analysis.This study evaluates the feasibility of using multimodal LLMs for detecting and recognizing military personnel in mobile device images. We compare the performance of Google’s Gemini 1.5 Pro, the open-source LLAVA model, and the newly released GPT-4o, analyzing their ability to identify military personnel and correctly classify attributes such as country and unit name. The dataset includes 436 images, comprising military personnel, military-themed mannequins, and civilians, to test the robustness of these models in real-world forensic scenarios.Our results show that all models achieved perfect precision (1.0) in detecting military presence, with high recall scores: 0.99 for Gemini, 0.98 for LLAVA, and 0.91 for GPT-4o. GPT-4o demonstrated a strong overall performance, with an accuracy of 0.751, placing it between Gemini (0.837) and LLAVA (0.69). However, contextual classification remains challenging. GPT-4o showed lower performance in country identification (0.385) compared to Gemini (0.787) but outperformed LLAVA (0.121). In unit name recognition, GPT-4o achieved 0.08, higher than LLAVA (0.005) but significantly lower than Gemini (0.254). All models struggled to distinguish mannequins from real military personnel, with high misclassification rates: 88 out of 99 mannequins were misclassified by Gemini, 86 by LLAVA, and comparable figures for GPT-4o.These findings highlight both the promise and current limitations of multimodal LLMs in forensic image analysis. While AI models can significantly assist forensic analysts in triaging large datasets and identifying relevant images, further research is needed to improve robustness against misclassification errors and enhance contextual understanding. This study provides valuable insights into the adoption of AI in digital forensics and lays the groundwork for future advancements in LLM-based forensic investigation techniques.