Detection-of-AI-generated-images

Reference paper - link

Datasets used:

• ArtiFact: Real and Fake Image Dataset
• AI recognition dataset

Note - All the mentioned datasets are partially used, training isn't done on all of their data.

Blog - link

An attempt to detect AI generated images in a generalized manner.

Logs:

Update: 17/02/2024

• Removed noise adding filters.
• experimented with usage of pixel_fluctuation_ratio as a feature to learn if image is ai generated or not. (refer to pixel_fluctuation.ipynb).

Update: 18/02/2024

• The standard matrix rotation function of scipy.ndimage.rotate uses affine transformation to rotate a matrix while there are more ways to rotate a matrix, attempt to explore methods to avoid noisy outputs after applying filters.

• Result - The way of rotation wasn't a problem but the way to merge filters was causing noisy outputs.

Update: 20/02/2024

• Implement data pipeline.
• Trained up till: loss: 0.1729 - binary_accuracy: 0.924.

Update: 21/02/2024

• implement testing notebook.

Update: 04/03/2024

• optimize for matrix rotation while applying filters.

Update: 12/03/2024

• Reimplement model architecture
• Models to be retrained

Update: 18/05/2024

• Fix filters issue
• Add preprocesseing demo