JPEGlitch is a specialized neural network designed to detect glitches in JPEG images. It combines the architectures of ResNet18 and SPPNet, providing the capability to process images of any size.

JPEGlitch is structured around an 18-layer Residual Network (ResNet) backbone, enhanced with a Spatial Pyramid Pooling (SPP) layer from SPPNet. This combination ensures that the model can handle input images of arbitrary sizes without the need for resizing or cropping, preserving the integrity and details of the original images. The network consists of approximately 11 million parameters, optimizing the balance between complexity and performance.

The model is trained on a custom dataset derived from the ImageNet-1k dataset.
This training set includes 2.46M samples, evenly divided into 1.23M positive examples (images without glitches) and 1.23M negative examples (images with glitches introduced through byte flipping).

JPEGlitch employs the AdamW optimizer. The learning rate is set to \(1 \times 10^{-4}\).

Accuracy ≈ 99.48%
Precision ≈ 99.32%
Recall ≈ 99.64%
F1 Score ≈ 99.48%