HYBRID ARCHITECTURES FOR FORENSIC DETECTION OF AI-GENERATED IMAGES:FROM LOW-RESOLUTION CLASSIFICATION TO ZERO-SHOT GENERALIZATION

Abstract

The rapid proliferation of generative AI models has fundamentally altered the land scape of digital media authenticity. Modern synthesis pipelines—spanning Generative Adversarial Networks, latent diffusion architectures, and large-scale text-conditioned models—can now produce images that are perceptually indistinguishable from gen uine photographs and artworks. This shift creates urgent demands for forensic detec tion systems that are simultaneously accurate, computationally tractable, and robust to generators not encountered during training. This thesis presents a progression of three studies that collectively address these demands through hybrid neural architectures operating in both spatial and frequency domains. The first study introduces HybridViT, a CNN–Vision Transformer framework in corporating Discrete Cosine Transform (DCT)-based frequency analysis for detecting AI-generated images in the low-resolution CIFAKE benchmark. By fusing local spa tial features extracted by convolutional layers, global context captured via multi-head self-attention, and high-frequency anomaly cues obtained through DCT projection, HybridViT achieves 94.95% test accuracy with only 75,009 trainable parameters— outperforming established CNN baselines including ResNet50, EfficientNet-B0, and VGG16 by substantial margins while requiring orders of magnitude fewer parameters. The second study extends this hybrid framework to the domain of fine art foren sics. ArtSpectraNet combines a lightweight CNN–Transformer spatial encoder with a Fast Fourier Transform (FFT)-based spectral branch for joint style classification and authenticity detection on the Kaggle AI ArtBench dataset. Evaluated on approximately 97,000 real and AI-generated artworks spanning ten historical movements, ArtSpec traNet achieves 99.61% authenticity detection accuracy alongside 73.18% top-1 style recognition—demonstrating that style-aware forensic reasoning is both feasible and necessary when the target domain is stylised art rather than natural photographs. The third and most technically demanding study addresses the critical problem of cross-generator generalisation. CONCORD (COnsistency-driven Network for CrOss scale Robust Detection) is a zero-shot AI-generated image detector grounded in two principles: learnable spectral basis projection and cross-scale structural consistency verification. The Adaptive Cosine Token Reconstruction (ACTR) module replaces fixed DCT projection with a learnable cosine basis bank whose spectral sensitivity adapts during training, while the Hierarchical Cross-Scale Token Consistency Atten tion (HCSTCA) mechanism verifies whether local spectral observations cohere with global image structure—a property natural images satisfy and synthetic ones system atically violate. All pretrained backbone encoders remain frozen, confining learning to the forensic reasoning modules. Trained on Stable Diffusion v1.4 alone and evalu ated zero-shot on seven unseen generators in the GenImage benchmark, CONCORD v achieves 78.1% average cross-generator accuracy, improving over prior methods by more than 3% overall and by more than 14 percentage points on challenging genera tors such as Midjourney and GLIDE. Taken together, the three studies advance understanding of how frequency-domain analysis, hybrid spatial–spectral architectures, and structural decoupling can be com bined to build AI-image forensics systems that are accurate, efficient, and generalis able.