[Paper Review] Block-level Double JPEG Compression Detection for Image Forgery Localization

Vinay Kumar Verma, Deepak Singh+1 more|arXiv (Cornell University)|Mar 20, 2020

Digital Media Forensic Detection4 citations

TL;DR

This paper proposes a deep learning-based method using DenseNet to detect single and double compressed JPEG blocks by combining quantized DCT coefficient histograms with corresponding quantization step sizes (q-factors), achieving state-of-the-art performance in forgery localization, especially for smaller blocks and unseen quantization matrices, without full decompression of the image.

ABSTRACT

Forged images have a ubiquitous presence in today's world due to ease of availability of image manipulation tools. In this letter, we propose a deep learning-based novel approach which utilizes the inherent relationship between DCT coefficient histograms and corresponding quantization step sizes to distinguish between original and forged regions in a JPEG image, based on the detection of single and double compressed blocks, without fully decompressing the JPEG image. We consider a diverse set of 1,120 quantization matrices collected in a recent study as compared to standard 100 quantization matrices for training, testing, and creating realistic forgeries. In particular, we carefully design the input to DenseNet with a specific combination of quantization step sizes and the respective histograms for a JPEG block. Using this input to learn the compression artifacts produces state-of-the-art results for the detection of single and double compressed blocks of sizes $256 imes 256$ and gives better results for smaller blocks of sizes $128 imes 128$ and $64 imes 64$. Consequently, improved forgery localization performances are obtained on realistic forged images. Also, in the case of test blocks compressed with completely different quantization matrices as compared to matrices used in training, the proposed method outperforms the current state-of-the-art.

Motivation & Objective

• Address the challenge of detecting forged regions in JPEG images by identifying single and double compressed blocks.
• Overcome the limitation of existing methods that rely only on standard 100 quantization matrices by using a diverse set of 1,120 real-world Q-matrices collected from a recent study.
• Improve forgery localization accuracy by learning compression artifacts directly from quantized DCT coefficients without dequantization.
• Enhance generalization to unseen quantization matrices, a critical requirement for real-world image forensics.
• Achieve superior performance on small block sizes (64×64, 128×128) and robust detection in realistic forgery scenarios.

Proposed method

• Extract quantized DCT coefficients directly from the JPEG bit-stream to avoid rounding and truncation errors from full decompression.
• Form histograms of the quantized DCT coefficients for each of the 64 frequency components in an 8×8 block.
• Use the corresponding quantization step sizes (q-factors) for each DCT frequency as a complementary input feature to the model.
• Design a specific input tensor combining the 64-bin histogram and the 64-element q-factor vector for each block to feed into DenseNet.
• Train a DenseNet architecture to classify each block as single or double compressed based on the joint representation of DCT histograms and q-factors.
• Apply a sliding window with stride 32 to predict double-compressed block probabilities across the image for forgery localization.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively distinguish between single and double compressed JPEG blocks using only quantized DCT coefficients and q-factors?
• RQ2 Does incorporating q-factors alongside DCT histograms improve detection performance compared to using histograms alone?
• RQ3 How does the proposed method generalize to unseen quantization matrices not present during training?
• RQ4 Can the method achieve better performance on smaller block sizes (64×64, 128×128) than existing state-of-the-art approaches?
• RQ5 To what extent does the method improve forgery localization accuracy on realistic forgeries involving copy-move and blur manipulations?

Key findings

• The proposed method achieved a 93.73% F1-score on 256×256 blocks and 91.26% on 128×128 blocks when using q-factors, significantly outperforming Park et al. [4] (90.30% and 87.68% respectively).
• For 64×64 blocks, the method achieved a 97.43% F1-score without q-factors and 95.68% with q-factors, demonstrating substantial improvement on smaller blocks.
• On copy-move forgery detection, the method achieved 79.92% F-measure, compared to 77.04% for Park et al. [4], showing improved localization performance.
• For blur manipulation, the method achieved 77.44% F-measure, up from 74.28% in Park et al. [4], indicating better detection of subtle manipulations.
• In the unseen Q-matrix scenario, the method achieved 92.83% test accuracy, outperforming Park et al. [4]’s 86.69%, demonstrating strong generalization capability.
• The model achieved 94.40% accuracy, 91.87% TPR, and 96.94% TNR on real-world forgery detection, surpassing Park et al. [4]’s 92.26%, 87.19%, and 97.33% respectively.