[Paper Review] Robust Contrast Enhancement Forensics Using Pixel and Histogram Domain CNNs

Pengpeng Yang, Rongrong Ni+3 more|arXiv (Cornell University)|Mar 13, 2018

Digital Media Forensic Detection12 references4 citations

TL;DR

This paper proposes two deep learning-based contrast enhancement (CE) forensics methods—P-CNN for the pixel domain and H-CNN for the histogram domain—using end-to-end convolutional neural networks to detect CE operations robustly. The methods achieve state-of-the-art performance in detecting CE without additional processing and show strong resilience against pre-JPEG compression and anti-forensic attacks.

ABSTRACT

Contrast enhancement (CE) forensics has always been ofconcern to image forensics community. It can provide aneffective tool for recovering image history and identifyingtampered images. Although several CE forensic algorithmshave been proposed, their robustness against some processingis still unsatisfactory, such as JPEG compression and anti-forensic attacks. In order to attenuate such deficiency, inthis paper we first present a discriminability analysis of CEforensics in pixel and gray level histogram domains. Then, insuch two domains, two end-to-end methods based on convo-lutional neural networks (P-CNN, H-CNN) are proposed toachieve robust CE forensics against pre-JPEG compressionand anti-forensics attacks. Experimental results show that theproposed methods achieve much better performance than thestate-of-the-art schemes for CE detection in the case of noother operation and comparable performance when pre-JPEGcompression and anti-foresics attacks is used.

Motivation & Objective

• Address the limited robustness of existing contrast enhancement forensics methods against pre-JPEG compression and anti-forensic attacks.
• Improve detection accuracy in the presence of image processing operations that obscure CE fingerprints.
• Develop end-to-end deep learning models capable of learning discriminative features directly from raw pixels and normalized gray-level histograms.
• Ensure stability across varying gamma correction levels and diverse image content.

Proposed method

• Proposes P-CNN, a pixel-domain CNN that applies a high-pass filter and batch normalization to reduce image content influence and balance data distribution.
• Designs a custom CNN architecture with empirically optimized width to enhance feature representation for CE detection in the pixel domain.
• Introduces H-CNN, a histogram-domain CNN that processes 256-dimensional normalized gray-level histograms as input to detect CE-induced peaks and gaps.
• Uses end-to-end training to learn discriminative features without handcrafted statistics, enabling robustness to preprocessing.
• Employs gamma correction modeling to analyze discriminability in pixel domain, with $ D_{\text{max}} $ derived as a function of $ \gamma $ to guide feature learning.
• Validates robustness through controlled experiments on JPEG-compressed images and anti-forensic attacks, using QF values (50, 70) and multiple anti-forensic techniques.

Experimental results

Research questions

• RQ1 Can deep learning models trained on pixel and histogram domains achieve superior robustness in detecting contrast-enhanced images compared to traditional handcrafted feature methods?
• RQ2 How do pre-JPEG compression and anti-forensic attacks affect the performance of existing CE forensics techniques?
• RQ3 To what extent can P-CNN and H-CNN maintain high detection accuracy under varying gamma correction levels and image processing conditions?
• RQ4 Does the scale of training data significantly impact the performance of P-CNN and H-CNN in CE forensics?

Key findings

• H-CNN achieves 99.90% detection accuracy for $ \gamma = 0.6 $ without any additional processing, outperforming the state-of-the-art Li[5] method (99.11%) under the same conditions.
• Under anti-forensic attack [14], H-CNN maintains 98.82% accuracy for $ \gamma = 0.6 $, while De Rosa[9] drops to 69.85% and Cao[4] fails entirely.
• For pre-JPEG compression at QF=50, both P-CNN and H-CNN achieve 99.9% accuracy, matching Li[5]’s performance, while De Rosa[9] drops to 70.26%.
• P-CNN shows superior robustness against anti-forensic attacks based on histogram manipulation, maintaining 98.6% accuracy for $ \gamma = 0.6 $, while H-CNN degrades significantly (88.77%) under the same attack.
• The performance of P-CNN improves with larger training data, demonstrating a positive correlation between data scale and detection accuracy.
• H-CNN is less sensitive to training data size due to its lower parameter count and compact input representation (256D histogram).