SMICA: Semantic and Multi-Granularity Integration with Constrained Adaptation for Fake Image Detection

The rapid development of generative and image-editing models has increased the prevalence of manipulated and synthetic images, creating substantial risks for security-sensitive applications. Most existing detectors target a single forensic domain and rely on domain-specific priors, limiting transfer to heterogeneous scenarios and unseen forgery types. We propose SMICA, a semantic and multi-granularity integration framework with constrained adaptation for fake image detection (FID). SMICA uses a frozen DINOv2 backbone with Low-Rank Adaptation (LoRA), enabling task-specific feature adaptation while updating only a small fraction of the model parameters. It further fuses the classification (CLS) token, Patch Mean, and Patch Max to capture global semantics, distributed local statistics, and sparse high-response cues. Adaptive Sharpness-Aware Minimization (ASAM) regularizes the trainable parameters toward flatter solutions during training. Under the unified OpenMMSec evaluation protocol, SMICA achieves AUC scores of 0.9608, 0.9685, 0.8809, and 0.8743 on Deepfake, AI-Generated Content (AIGC), Image Manipulation Detection (IMD), and Document Image Forgery Detection (Doc), respectively, yielding a macro-average AUC of 0.9211 while optimizing only 2.50% of the total parameters. These results show that combining multi-granularity forensic evidence with parameter-efficient constrained adaptation provides an effective solution for generalizable FID.