FMACNN: Federated multi-attention CNN framework for artificial image detection

초록

The rapid advancement of generative artificial intelligence has facilitated the creation of photorealistic images, raising concerns about the spread of misinformation. Detecting artificially generated fake images has become crucial due to their potential impact on public opinion and various sectors. Additionally, ensuring the detection of these synthetic images while preserving users' privacy is equally necessary. Privacy-preserving detection methods are essential to protecting user confidentiality, preventing unauthorized access to sensitive data, and combating the spread of deceptive visual content. This study proposes a novel Federated Multi-Attention CNN (FMACNN) framework that detects artificially generated images on edge devices. This approach combines the strengths of multi-attention mechanisms with the privacy-preserving, decentralized nature of federated learning (FL), thereby ensuring a comprehensive and robust detection system. This research also introduces the Realistic AI-Generated Image (RealAIGI) dataset, which significantly advances Artificial Image Detection (AID) by providing a diverse set of AI-generated images. Leveraging the cutting-edge Diffusion Transformer model (Sora), the dataset provides challenging scenarios for enhanced training and real-world applicability. The proposed model achieves 99.12% accuracy, outperforming current state-of-the-art models on the RealAIGI dataset and surpassing benchmarks CIFAKE and FaceForensics++. Likewise, the proposed FL framework demonstrated greater scalability and resource efficiency than SOTA frameworks. By leveraging the proposed framework and the RealAIGI dataset, this research aims to advance the state of the art in artificial image detection and to provide a robust solution.

키워드