CAIAMAR: A Multi-Agent Reasoning-Driven Context-Aware Image Anonymization Framework

CAIAMAR is a context-aware image anonymization framework based on multi-agent reasoning. Through a three-agent PDCA cycle coordination mechanism, it combines spatial context to determine PII types, reducing re-identification risk by 73% on the CUHK03-NP dataset while maintaining image quality and semantic segmentation integrity. This framework addresses the issues of over-processing/under-processing in traditional anonymization methods and data sovereignty, opening up new directions in the field of privacy computing.

Street view images contain a large amount of personally identifiable information (PII), but their identification is highly context-dependent. Traditional anonymization faces a dilemma: over-processing impairs image usability, while under-processing misses indirect identifiers; API-based solutions expose data, violating the principle of data sovereignty. Existing computer vision (CV) methods use rigid category rules and cannot distinguish the privacy sensitivity of the same object in private/public spaces, making spatial context understanding a key research topic.

CAIAMAR adopts three-agent PDCA cycle collaboration: the Reconnaissance Agent uses a "reconnaissance-zoom" strategy to coarsely locate potential sensitive areas; the Segmentation Agent performs open-vocabulary local segmentation; the Deduplication Agent detects duplicate targets based on a 30% IoU threshold. The architecture leverages the reasoning capabilities of Large Vision-Language Models (LVLM) to determine PII based on spatial context rather than fixed category rules.

The core innovation is the spatial filtering coarse-to-fine strategy: first determine whether the area belongs to private territory or public space to decide the anonymization intensity. It uses modality-specific diffusion guidance to reduce re-identification risk through appearance decorrelation while preserving semantic consistency. The framework runs entirely locally (using open-source models), generates human-readable audit trails, and supports GDPR compliance.

1. Re-identification risk: The R1 metric on the CUHK03-NP dataset decreased from 62.4% to 16.9% (a 73% reduction), effectively handling indirect PII such as clothing and accessories; 2. Image quality: On the CityScapes dataset, KID=0.001 and FID=9.1, outperforming existing methods; 3. Downstream compatibility: The processed images have good semantic segmentation performance and retain the features required for scene understanding.

Contributions: 1. The agent workflow enables open-source models to surpass the robustness of proprietary models; 2. Full local operation ensures data sovereignty, and audit trails meet compliance requirements. In the future, agents for time-series analysis and multi-modal fusion can be expanded to improve the context understanding accuracy of LVLMs. This research promotes privacy protection from a "one-size-fits-all" approach to context-aware, and from black-box to transparent systems.