Neural network based block-level detection of same quality factor double JPEG compression

Detecting double JPEG compression with the same quality factor is an open research problem in image forensics. The subtle artifacts of double JPEG compression with the same quantization matrices are often indistinguishable by traditional approaches such as histogram-based feature extraction. Existing approaches fail to effectively classify smaller size patches, due to the absence of sufficient information at smaller scales. To tackle this, we propose a new feature based on the difference of quantized DCT coefficients, which is relatively invariant to the size of the image or image patch. For classification, we utilize the multi-layer-perceptron (MLP) network. We compare our results on multiple patch sizes and quality factors, on the UCID dataset with the existing approaches. Using our proposed feature in conjugation with the existing feature vector along with the use of MLP, we observed a maximum of 1.52% increase in accuracy for smaller sized patches (128 × 128), compressed with a quality factor of 60.