Convolutional Neural Network for Image Forgery Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional image forgery detection techniques are limited in their ability to accurately detect manipulations in compressed images and are often specific to certain file formats and environments, making them ineffective in diverse operations and formats, especially after double JPEG compression which introduces additional data loss.

Innovation Solution

A convolutional neural network system that includes a manipulated feature pre-processing unit, extraction unit, refining unit, and classification unit, utilizing a two-stream neural network structure with constrained convolutional layers and Markov statistics-based networks to detect image manipulation in JPEG-compressed images, capable of distinguishing between normal and manipulated image blocks, and producing a forgery confirmation map.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional image forensic techniques are used to detect image forgery, then detection can be performed in specific environments, but the techniques are limited to specific file formats and operations, reducing their adaptability

Engineering Contradiction:

Improvedetection accuracyVSAvoidapplicability to diverse formats and operations

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by developing a detection system based on convolutional neural networks that can handle multiple file formats (JPEG, PNG, BMP) and various manipulation operations (copy-move, splicing, filtering) through a unified architecture. The system processes images by converting them to a standardized internal representation, allowing the same detection algorithms to work across different formats and manipulation types without requiring format-specific or operation-specific detection modules.

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Productivity

If JPEG compression is applied to images, then image storage and transmission are facilitated, but data loss occurs making it difficult to track manipulation traces

Engineering Contradiction:

Improveimage processing efficiencyVSAvoidmanipulation trace data

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The patent converts the harmful effect of JPEG compression (data loss) into a beneficial feature by training the convolutional neural network to recognize manipulation traces that persist even after compression. The system learns to detect subtle patterns and artifacts that remain after JPEG encoding, transforming the information loss into a detection advantage by focusing on the residual traces that compression fails to eliminate completely.

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

3Ease of manufacture

If double JPEG compression is applied to manipulated images, then image editing and re-saving are completed, but additional data loss occurs reducing detection capability

Engineering Contradiction:

Improveimage editing convenienceVSAvoidmanipulation detection accuracy

Core Design Contradiction:

Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by pre-training the convolutional neural network on extensively compressed images, including those that have undergone multiple JPEG encoding cycles. The system learns the characteristic patterns and artifacts that result from repeated compression, enabling it to detect manipulations even when the image has been saved and re-loaded multiple times. This pre-training ensures the detector is prepared for the degraded quality conditions that result from double or multiple compression.

Inventive Principle:

Principle #10Preliminary action

4Ease of operation

If sophisticated image editing software is used, then image forgery can be easily created, but the complexity of detecting such forgery increases

Engineering Contradiction:

Improveimage editing accessibilityVSAvoiddetection system complexity

Core Design Contradiction:

Ease of operationVSDevice complexity

Solution Approach 1:

The patent replaces traditional mechanical or rule-based detection systems with a data-driven convolutional neural network approach. Instead of relying on hand-crafted features and explicit detection rules that require complex conditional logic, the system uses learned representations from training data to automatically identify manipulation patterns. This substitution simplifies the detection architecture while improving its ability to handle diverse and sophisticated forgeries created by various editing software.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS11861816B2System and method for detecting image forgery through convolutional neural network and method for providing non-manipulation detection service using the same

Publication Date: 2024.01.02 NHN CORP

AI summary

A system and a method for detecting image forgery through a convolutional neural network are capable of detecting image manipulation of compressed and/or color images. The system comprises a manipulated feature pre-processing unit applying an input image to a high-pass filter to enhance features due to image forgery; a manipulated feature extraction unit extracting image manipulated feature information from the image with the enhanced features through a pre-trained convolutional neural network; a feature refining unit refining the extracted image manipulated feature information; and a manipulation classifying unit determining the image forgery based on the image manipulated feature information refined by the feature refining unit.