| dc.description.abstract |
The automatic detection of AI-generated art images is essential for
distinguishing authentic human creations from artificial ones. This process is critical
for authenticity verification, provenance control, misinformation management, and
digital forensics. With the rapid evolution of deep learning content generation, the
existing detection approaches within artistic imagery remain an underexplored domain
characterized by artworks that differ widely in style and often contain non-standard,
complex, or distorted visual patterns. The proposed system is an efficient deep learning
framework for deepfake detection on art image data. It is a fine-tuned CNN-based
generative art detection to classify real and human-created art accurately by learning
discriminative visual features such as texture, structure, and statistical patterns,
adapting a pretrained CNN model and also finetuning architecture layers and defining
the spatial dimension, which is used to determine the level of detail captured in feature
extraction and classification. The framework evaluates pretrained VGG16,
DenseNet121, ResNet50, and EfficientNetB0 architectures on a balanced dataset
comprising authentic and multi-model-generated art images with stylistic diversity. The
models are optimized through architecture fine-tuning, task-specific input
representation, spatial dimension selection, hyperparameter tuning, data augmentation,
and layer optimization to enhance discriminative feature extraction and classification.
Their performance is benchmarked against a traditional baseline to assess accuracy,
adaptability, and generalization. The findings demonstrate that transfer learning,
combined with dataset diversity and augmentation, enhances generalization and
provides an effective framework for art deepfake detection, offering practical
implications for preserving artistic authenticity in an era of increasingly sophisticated
AI-generated art imagery. |
en_US |