ICACT20250214 Slide.23
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
References(Continued…):
|
|
ICACT20250214 Slide.22
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
References(Continued…):
|
|
ICACT20250214 Slide.21
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
References(Continued…):
|
|
ICACT20250214 Slide.20
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
References:
[1] Hany Farid. Exposing digital forgeries from jpeg ghosts. IEEE transactions on information forensics and security, 4(1):154–160, 2009.
[2] Alessandro Piva. An overview on image forensics. International Scholarly Research Notices, 2013(1):496701, 2013.
[3] Jessica Fridrich. Digital image forensics. IEEE Signal Processing Magazine, 26(2):26–37, 2009.
[4] Thomas Huang, GJTGY Yang, and Greory Tang. A fast two-dimensional median filtering algorithm. IEEE transactions on acoustics, speech, and signal processing, 27(1):13–18, 1979.
[5] Jaakko Astola and Pauli Kuosmanen. Fundamentals of nonlinear digital f iltering. CRC press, 2020.
|
|
ICACT20250214 Slide.19
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
REFERENCES
|
|
ICACT20250214 Slide.18
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Future Work
Future Work:
Enhancing Model Generalization:
Train on diverse datasets with different types of image manipulations.
Explore domain adaptation techniques for better performance on unseen data.
Extending to Other Manipulations:
Detecting other types of forgeries such as splicing, copy-move, and inpainting.
Apply SobelMNet to video forensics for detecting frame manipulations.
Improving Robustness:
Investigate the model's behavior under extreme compression and low-quality images.
Implement adversarial training to make the model resistant to sophisticated tampering techniques.
Real-World Deployment:
Optimize for mobile and embedded devices for on-the-go forensic analysis.
Develop a real-time image forensics tool integrating SobelMNet for law enforcement and cybersecurity.
|
|
ICACT20250214 Slide.17
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Conclusion
SobelMNet effectively detects median filtering with 99.43% accuracy on compressed images.
Outperforms existing methods like MedFCN and Streak Analysis.
Uses Sobel filter preprocessing to enhance detection accuracy.
Lightweight and efficient model, suitable for forensic applications.
Robust performance across multiple manipulations, including Gaussian Blur, Resampling, and AWGN.
|
|
ICACT20250214 Slide.16
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Comparative Analysis (Extended)
Performance Against Existing Methods:
SobelMNet outperforms MedFCN and Streak Analysis in detecting median filtering, even under compressed settings.
Handles low-resolution images better than traditional deep learning models.
Faster inference time due to its lightweight CNN structure.
Robustness to Multiple Manipulations:
Detects median filtering even when combined with Gaussian Blur, Resampling, and Additive White Gaussian Noise (AWGN).
Achieves 100% accuracy when detecting median filtering + AWGN.
Maintains high accuracy across different levels of compression and scaling.
Computational Efficiency:
Uses fewer parameters than conventional CNN models.
Requires less training data compared to deeper architectures.
Can be implemented in real-time forensic applications.
Limitations:
Slightly lower accuracy when detecting median filtering + resampling due to structural distortions.
Performance may degrade with extreme compression (QF < 50).
|
|
ICACT20250214 Slide.15
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Comparative Analysis
|
|
ICACT20250214 Slide.14
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Experimental Results
Binary Classification Performance:
Training Accuracy: 96.83%
Validation Accuracy: 98.89%
Median Filtering Detection Accuracy: 99.43%
Multi-Class Classification Performance:
Best Validation Accuracy: 98.32%
Handles multiple manipulations effectively.
Comparison with Existing Models:
Outperforms MedFCN and Streak Analysis in accuracy and robustness.
|
|
ICACT20250214 Slide.13
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Dataset & Experimentation
Dataset: Dresden Image Dataset (10,000 images).
Preprocessed as 64×64 grayscale images.
Manipulation types applied:
Median Filtering (3×3, 5×5)
Gaussian Blurring
Resampling
Additive White Gaussian Noise (AWGN)
Train-Test Split:
30,000 training images
6,000 validation images
10,000 test images
|
|
ICACT20250214 Slide.12
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
SobelMNet Architecture
Three main blocks:
Preprocessing Block: Sobel filter enhances image edges.
Feature Extraction Block:
Four convolutional layers with batch normalization and ReLU activation.
Max-pooling for dimensionality reduction.
Classification Block:
Fully connected layers for binary/multi-class classification.
Adam optimizer and CrossEntropyLoss for training.
|
|
ICACT20250214 Slide.11
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Preprocessing with Sobel Filter
Three main blocks:
Preprocessing Block: Sobel filter enhances image edges.
Feature Extraction Block:
Four convolutional layers with batch normalization and ReLU activation.
Max-pooling for dimensionality reduction.
Classification Block:
Fully connected layers for binary/multi-class classification.
Adam optimizer and CrossEntropyLoss for training.
|
|
ICACT20250214 Slide.10
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
PROPOSED WORK: CNN-BASED MODEL (Extended)
|
|
ICACT20250214 Slide.09
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
PROPOSED WORK: CNN-BASED MODEL
|
|
ICACT20250214 Slide.08
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Proposed Methodology - SobelMNet
A lightweight CNN-based model for detecting median filtering in compressed images.
Key Features:
Sobel filter-based preprocessing: Enhances edge residuals for better feature extraction.
Feature extraction using CNN: Extracts structural differences between original and manipulated images.
Lightweight architecture: Efficient model for both binary and multi-class classification.
Robust evaluation: Tested on low-resolution, compressed images for real-world forensic applications.
|
|
ICACT20250214 Slide.07
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Research Gaps
Reduced accuracy when median filtering is combined with other operations.
High dependency on high-resolution images, making detection difficult for resized/compressed images.
Existing CNN models are computationally expensive.
Need for a lightweight, robust model to handle multiple manipulations effectively.
|
|
ICACT20250214 Slide.06
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Related Work
|
|
ICACT20250214 Slide.05
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Introduction
Digital images are extensively used in social media, legal evidence, and medical fields.
Image manipulation has become easier with advanced tools.
Median filtering is a common method to remove noise but can be misused for forgery.
Detecting median filtering is crucial for forensic analysis.
Deep learning models have shown promise in forensic tasks but struggle with compressed and low-resolution images.
|
|
ICACT20250214 Slide.04
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Abstract
Digital image forgery is a major concern, requiring effective detection methods.
Median filtering is widely used to smooth images but can obscure tampering.
SobelMNet is a lightweight CNN designed to detect median filtering in compressed images.
Utilizes Sobel filter-based preprocessing to enhance residual differences.
Achieves 99.43% accuracy in detection and outperforms state-of-the-art models.
|
|
ICACT20250214 Slide.03
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Anukriti Kaushal1, Rajeev Kumar1, Sanjay Kumar1
1Dept. of Computer Science & Engineering, Delhi Technological University, Delhi, India
anukritikaushal@dtu.ac.in, rajeevkumar@dtu.ac.in, sanjay.kumar@dtu.ac.in
|
|
ICACT20250214 Slide.02
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
SobelMNet: A Lightweight CNN based Approach
for Median Filtering Detection in compressed images
|
|
ICACT20250214 Slide.01
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
The 27th International Conference on Advanced Communication Technology, ICACT 2025
|
