AgentForensics: An Open-Protocol Multi-Agent Framework for Synthetic Image Analysis

Authors:
Paolo Bestagini*, Mohamed Aziz Cherif*, Ashkriti Dewan*, Daniele Ugo Leonzio^, Sara Mandelli*, Daniele Moro^, Davide Redaelli*, Vivek Kumar Sharma*

Affiliations:
* DEIB – Politecnico di Milano – Milan, Italy
^ Realynx - Milan, Italy

Published at: Deep Learning and Mathematical Methods for Deepfake Detection Workshop at ICANN 2026

Citation:

@inproceedings{Bestagini2026AgentForensics,
  title={AgentForensics: An Open-Protocol Multi-Agent Framework for Synthetic Image Analysis},
  author={Bestagini, Paolo and Cherif, Mohamed Aziz and Dewan, Ashkriti and Leonzio, Daniele Ugo and Mandelli, Sara and Moro, Daniele and Redaelli, Davide and Sharma, Vivek Kumar},
  booktitle={Artificial Neural Networks and Machine Learning. ICANN 2026 International Workshops},
  year={2026}
}

Paper: Full paper PDF

Abstract

Synthetic image forensic analysis is increasingly difficult to address with single-purpose forensic detectors, since detectors trained on one generator family, post-processing pipeline, or evidence modality may fail under distribution shift. This paper presents a preliminary system description of AgentForensics, a multi-agent framework for synthetic image forensic analysis in which independently deployable detector agents are coordinated by an LLM-based orchestrator. Detector agents expose a uniform interface returning a synthetic-origin probability, an abstention flag, and optional natural-language evidence. The orchestrator combines these heterogeneous outputs into a structured verdict, using a deterministic numeric aggregation rule as a reference while allowing qualitative evidence to adjust the final decision when explicitly justified. The framework is extensible and uses two open protocols. Agent2Agent (A2A) supports runtime automatic discovery and invocation of detector agents, while the Model Context Protocol (MCP) decouples each detector agent from its underlying inference backend. As a proof of concept, we implement the framework with three complementary modalities: patch- based low-level statistics for synthetic image detection, data-driven forgery localization, and metadata-based provenance analysis. A worked example illustrates how conflicting detector outputs and qualitative reports interact in an auditable execution trace. The system is functional but not yet benchmarked; we discuss current limitations and the main open questions rather than presenting final results.