A Comprehensive Survey on Deep Learning Techniques for Image Forgery Detection with Real-

Time Forensic Applications

Pranav S Bharadwaj, Don Bosco Institute of Technology, Bangalore

Sinchana K P, Don Bosco Institute of Technology, Bangalore

Rakshith M L, Don Bosco Institute of Technology, Bangalore

R A Abhishek, Don Bosco Institute of Technology, Bangalore

Raghunath, Don Bosco Institute of Technology, Bangalore

Abstract

The growing sophistication of image manipulation methods, ranging from straightforward splicing
and copy-move actions to highly realistic GAN-generated content, presents significant challenges
for digital media forensics. In response, deep learning has emerged as a leading approach for
detecting image forgeries, leveraging hierarchical feature learning and end-to-end training for
improved robustness and generalization. This paper takes a close look at the best ways deep
learning can help us find fake images. We cover different types of computer models, like those
using CNNs, attention mechanisms, and transformers. We've also come up with a new, simple way
to sort these methods based on: what kind of trickery was used, how detailed the fake is, if it can
work instantly, and how easy it is to understand for legal reasons. We also talk about attacks that
try to fool these detection systems and ways to fight back, explaining how these things affect how
reliable the detection is. A big problem right now is that the datasets used for testing often don't
have the right legal permissions and don't show all the many ways images can be faked. This clearly
means we really need new datasets that are both legally approved and full of all sorts of fake
images.

Keywords

Image Forgery Detection, Deep Learning, Digital Forensics, Real-Time Detection, CNNs, GAN-
based Forgeries, Model Interpretability, Adversarial Attacks

Introduction

In the digital age, images serve as vital evidence in journalism, legal proceedings, intelligence
operations, and social media platforms. It's become super easy to change digital pictures and
videos, which makes us seriously wonder if we can really trust what we see. Older ways of faking
images, like copying parts, sticking different pictures together, or simply touching them up, used
to be caught by older methods. But now, with new tools like GANs (which are like smart faking
machines), neural rendering, and diffusion models, people can create incredibly real-looking
fake images. Even skilled humans have challenges identifying these fakes, let alone the outdated
tools we used to check for them.

Our earlier ways of identifying them, which primarily depended on minor details, camera noise, or
specifics in JPEG files, are now falling behind due to the vast quantity of these devious fakes.
Although these old techniques had been useful in the past, they frequently break down when images
are altered (e.g., reduced in size or given additional noise) or intentionally altered. Also, these older
methods are less effective for all the various types of fakes because they rely on features that must
be manually configured.

Researchers are now using deep learning to identify fake images in an effort to address this
expanding issue. Important details can be learned directly from the image data itself with the help
of deep learning systems, particularly Convolutional Neural Networks (CNNs), Vision
Transformers (ViTs), and mixed approaches. When it comes to determining the exact location of
image alteration, identifying the type of fake, checking whether the image makes visual sense, and
even identifying fakes in videos, these more recent models outperform the outdated methods. Deep
learning models are able to detect subtle, hidden gaps and can even adapt to new kinds of fakes that
they haven't previously came across, unlike the more traditional, manual approaches.

Even all of this advancement, there is still no universally accepted understanding of how to use
deep learning to detect fakes in a way that is suitable for legal use. This is particularly true for tools
that have to work quickly, use little computer power, and be legal. Currently, the majority of
research focuses on how well these intelligent computer programs function on test sets, frequently
ignoring the needs of the real world. These requirements include being able to demonstrate
precisely how a fake was found, describing how the detection process works, how fast it detects
fakes, whether it can operate on small devices like phones, and how well it survives attacks
designed to fool it.

This paper seeks to address this gap by presenting a comprehensive, technical, and forensic-
oriented survey of deep learning techniques for image forgery detection. We extend beyond simple
architectural taxonomy, introducing a novel classification framework based on manipulation type,
network design, detection granularity (image-level vs. pixel-level), and deployment feasibility
(cloud-based, mobile, edge-compatible). Additionally, we assess forensic readiness by examining
criteria such as model explainability (e.g., saliency, attention heatmaps), reproducibility under legal
scrutiny, and resistance to adversarial perturbations. We further outline the outstanding challenges
that impede the real-world adoption of forgery detection models, such as detecting low-resolution
or heavily compressed images, operating in real-time forensic workflows, and ensuring evidential
integrity in judicial contexts.

1.Literature Review

[1] Li Jiang, 2024

The research paper titled "An Effective Image Copy-Move Forgery Detection Using Entropy
Information" introduces a novel and efficient framework that enhances traditional keypoint-based
image forgery detection by integrating entropy information into the Scale Invariant Feature
Transform (SIFT) process. This innovation significantly improves the detection of tampered
regions, particularly in smooth or low-texture areas where conventional methods often fall short.
By incorporating entropy images to guide keypoint localization and employing an overlapped
entropy level clustering algorithm, the proposed method achieves high detection accuracy while
maintaining a reasonable computational load. Extensive experiments on benchmark datasets like
GRIP and CMH demonstrate that this approach outperforms several existing algorithms in terms
of both image-level and pixel-level detection metrics. It successfully balances performance and
time efficiency, making it practical for real-time forensic applications.

[2] Abeer Oraby, Ayman El-Sayed and Ezz El-Din Hemdan1, 2024

The research paper titled "An Efficient Image Forgery Detection Framework using Transfer
Learning Models" proposes a powerful deep learning-based approach for detecting forged images,
utilizing pre-trained convolutional neural networks (CNNs) through transfer learning. The
framework includes various models such as ResNet50, VGG16, Xception, DenseNet201,
MobileNet, and NasNetMobile, achieving its best performance with ResNet50, which yielded a
remarkable 96% accuracy on the MICC-F220 dataset. This approach is particularly effective due
to its ability to harness the feature extraction capabilities of established CNN architectures,
enabling accurate classification of authentic versus tampered images. The system is designed for
real-time application, supported by comprehensive preprocessing steps that optimize performance
and minimize computational overhead.

[3] ANJALI DIWAN AND ANIL K. ROY, 2024

The paper titled "CNN-Keypoint Based Two-Stage Hybrid Approach for Copy-Move Forgery
Detection" introduces a robust and innovative method for detecting and localizing copy-move
forgery by integrating CenSurE keypoint detection with a CNN-based deep learning model. This
two-stage hybrid framework leverages the precision of keypoint matching and the learning
capabilities of convolutional networks, improving detection even under complex post-processing
conditions like JPEG compression, noise addition, geometric transformations, and brightness or
contrast variations. The model has been tested across diverse datasets such as CMFD, CoMoFoD,
and CASIA-II, demonstrating superior performance in terms of accuracy and processing time
compared to several state-of-the-art methods. Its design supports the detection of forgeries in both
low-texture and highly textured images, and it excels in identifying multiple and geometrically
transformed forgeries, making it a powerful tool for real-time forensic applications.

[4] Nandini Kashyap, Prince Yadav, Nikita and Anukriti Kaushal 2024

The paper titled "Deep Learning Strategies for Effective Image Forgery Detection and
Localization" presents an advanced and efficient approach to image forgery detection using two
hybrid deep learning models: VGG19+U-Net and ResNet34+LinkNetThe problems of accurately
finding and splitting forged regions—particularly when it comes to splicing and copy/move
manipulations—are specifically addressed by these models. The suggested models effectively and
highly correctly identified tampered areas by fusing the spatial preservation of U-Net and LinkNet
with the deep feature extraction powers of VGG19 and ResNet34. Experiments on the IEEE IFS-
TC Image Forensics Challenge dataset show that the ResNet34+LinkNet model works better than
traditional methods, with an F1 score of 0.9557. For this, it works effectively in forensic
applications in the real world where precision in localization and accuracy are crucial. Still, the
suggested approach has limitations despite its great results.
[5] Chandana S, Nagarathna CR, Amrutha A, Jayasri A 2024

The research paper titled "Detection of Image Forgery Using Error Level Analysis" proposes an
effective approach for identifying manipulated images through the integration of Error Level
Analysis (ELA) and Convolutional Neural Networks (CNN). ELA is employed to detect regions
in an image with inconsistent compression levels—an indicator of tampering—while CNNs are
utilized for their powerful feature extraction and classification capabilities. The model, trained on
the CASIA v2 dataset, achieves strong results with a training accuracy of 97.18% and a test
accuracy of 87.75%, making it a promising tool for practical forgery detection. This method not
only detects whether an image is forged but also highlights the forged regions, enhancing its utility
for forensic applications.

[6] Ayesh Meepaganithage, Suman Rath, Mircea Nicolescu and Monica Nicolescu

The research paper titled "Image Forgery Detection Using Convolutional Neural Networks"
investigates the effectiveness of CNN architectures such as ResNet and VGG in identifying forged
images. Among the evaluated models, ResNet-101 demonstrated the best performance, achieving
93.46% accuracy, 95.08% precision, and a 93.57% F1-score, successfully distinguishing both
authentic and manipulated images. The study emphasizes the capability of CNNs to learn subtle
image inconsistencies such as texture, lighting, and noise patterns, which are vital for identifying
cloning, splicing, retouching, and deepfake manipulations. Through rigorous experimentation
using a balanced Kaggle dataset of 13,000 images, the authors show that CNN-based detection is
both accurate and computationally efficient when employing pre-trained models and transfer
learning.

2.Comparative Analysis

Some research ([1], [5]) uses simple methods like looking for image errors, which are easy but
struggle with complex fakes. Other studies ([2], [3], [4], [6]) use deep learning, which is much
more accurate for tricky fakes but needs powerful computers and lots of data. For example, some
work ([1], [3]) focuses on finding copied image parts, with [3] using a smart hybrid model. Paper
[2] shows how using already-trained models saves time and boosts accuracy, especially with
limited data, while [4] helps pinpoint exact changes for investigations.

Author(s) Year Methodology Advantages Disadvantages Results

L. Jiang & Z. 2024 Entropy-based Good for High High F1 and TP

Lu keypoint smooth region computational rates on public
detection with detection; complexity; datasets
clustering improved lacks semantic
accuracy analysis
A. Oraby & A. 2022 Transfer Fast, high No tamper 96% accuracy
El-Sayed learning with accuracy, localization; (ResNet50 on
CNNs minimal depends on MICC-F220)
(ResNet50, training pretrained
VGG16, etc.) required networks

A. Diwan & A. 2024 Hybrid CNN- Robust against Complex High accuracy
K. Roy keypoint with postprocessing; architecture; on CMFD,
CenSurE and good higher resource CoMoFoD,
deep model localization usage CASIA-II

N. Kashyap et 2024 VGG19+U-Net Precise High resource F1 score: 0.9557

al. and localization usage; needs (ResNet+LinkNe
ResNet34+Link and annotated t)
Net segmentation masks

Chandana S et 2024 ELA + CNN Interpretable Not robust for 87.75% test
al. ; efficient on uniformly accuracy
ELA- compressed
prepared data images

A. 2024 CNNs Best VGG 93.46% accuracy

Meepaganithag (ResNet101, performance underperformed; (ResNet101)
e et al. VGG) with no localization
ResNet101;
transfer
learning
Sharma & 2023 Vision Strong for Heavy model; Significant
Bhargava Transformer splicing requires large improvement in
with patch-wise detection; datasets splicing
training captures accuracy
global
context

Jadhav & Pawar 2023 Deep Residual Good for Needs GPU; Improved
Networks compressed limited to detection under
image copymove compression
detection

Bhattacharya & 2023 CNN + Effective Complex High F1 score on

Mandal Attention feature model; may multimedia
mechanism focus; good overfit small datasets
localization datasets

Kumar & Singh 2023 Multi-scale Robust Complex Effective on

deep splicing design; slower forensic splicing
supervision + detection; inference datasets
consistency consistent
check output

Ahmed & 2023 Frequencyaware Specialized Less High accuracy for

Jamal CNN for GAN generalizable to GAN-
image traditional generated
detection forgery images

Wang, Li & 2023 Dense feature Efficient Needs careful High precision in
Zhang fusion CNN deep model tuning; copy-move
with fused computational detection
features cost
 Chen, Kang & 2021 Two-stream Good Dual-stream Strong results on
Bui CNN (color + forgery complexity; localization tasks
noise streams) localization requires
preprocessing

Wang, Wang & 2021 Harmonic Attention May require Effective on

Zhang attention improves fine-tuning per various forgery
network feature dataset types
extraction

Tran & Choi 2021 Lightweight Low Limited Good trade-off

CNN + block resource generalization to between speed
matching usage; complex and accuracy
interpretable forgeries

Conclusion

In this survey, we critically examined the landscape of deep learning techniques applied to image
forgery detection, focusing on their applicability in real-time and forensic contexts. The study
highlights the increasing dominance of deep neural networks, particularly CNNs, hybrid
architectures, and transformer-based models, in identifying various types of image manipulations,
from classic techniques like splicing and copy-move forgeries to more advanced GANs. While
these methods demonstrate high detection accuracy and promising localization capabilities, they
often fall short in generalization, clarity, and resistance to adversarial manipulation. Additionally,
real-time deployment remains a challenge due to computational limitations and the absence of
legally curated, manipulation-diverse datasets. Our analysis also underscores the need for
explainable AI to ensure forensic readiness and legal admissibility. Moving forward, research
should prioritize the development of lightweight, explainable, and robust models, supported by
comprehensive datasets and enhanced by technologies like blockchain for traceability.

References

[1] L. Jiang and Z. Lu, "An Effective Image Copy Move Forgery Detection Using Entropy
Information," in 2024 9th International Conference on Image, Vision and Computing (ICIVC),
2024, pp. 179-186.

[2] A. Oraby and A. El-Sayed, "An Efficient Image Forgery Detection Framework using
Transfer Learning Models," International Journal of Computer Vision and Image Processing
(IJCVIP), vol. 12, no. 1, pp. 12-22, 2022.
[3] A. Diwan and A. K. Roy, "CNN-Keypoint Based TwoStage Hybrid Approach Copy-Move
Forgery for Detection," IEEE Access, vol. 12, pp. 43809-43824, 2024.

[4] N. Kashyap, P. Yadav, Nikita, and A. Kaushal, "Deep Learning Strategies for Effective
Image Forgery Detection and Localization," in 2024 International Conference on Computer,
Electronics, Electrical Engineering & their Applications (IC2E3), 2024, pp. 490 495.

[5] Chandana S, Jayasri A, and Nagarathna C R, "Detection Of Image Forgery Using Error
Level Analysis," in 2024 International Conference on Intelligent and Innovative Technologies in
Computing, Electrical and Electronics (IITCEE), 2024,1 pp. 649-653.

[6] A. Meepaganithage, M. Nicolescu, S. Rath, M. Nicolescu, and S. Sengupta, "Image Forgery

Detection Using Convolutional Neural Networks," in 2024 12th International Symposium on
Digital Forensics and Security (ISDFS), 2024, pp. 1-6.

[7] Sharma, K., & Bhargava, D. (2023). "Enhanced Splicing Detection using Vision
Transformer with Patch wise Training." Pattern Recognition Letters, 169, 80-88.

[8] Jadhav, A., & Pawar, V. (2023). "Forgery Detection in Compressed Images Using Deep
Residual Networks." Signal, Image and Video Processing, 17(2), 317-327.

[9] Bhattacharya, D., & Mandal, B. (2023). "Copy Move Forgery Detection Using CNN and
Attention Mechanisms." IEEE Transactions on Multimedia, 25, 4863-4876.

[10] Kumar, A., & Singh, R. (2023). "Robust Splicing Forgery Detection Using Multi-Scale
Deep Supervision and Consistency Check." Forensic Science International: Digital Investigation,
44, 301556.

[11] Ahmed, W., & Jamal, T. (2023). "Detecting GAN-Generated Images Using Frequency-
aware CNN Models." Neural Computing and Applications, 35(11), 8695-8706.

[12] Wang, R., Li, H., & Zhang, J. (2023). "An Efficient Deep Learning Model for Copy-Move
Forgery Detection Based on Dense Feature Fusion." IEEE Access, 11, 21732-21744.

[13] Chen, J., Kang, J., & Bui, T. (2021). Two-stream networks for image forgery localization.
IEEE Access, 9, 3185-3197.

[14] Wang, S., Wang, Y., & Zhang, H. (2021). Harmonic attention network for forgery detection.
IEEE Transactions on Image Processing, 30, 4060-4072.

[15] Tran, D. Q., & Choi, Y. (2021). Image Forgery Detection Based on Lightweight CNN with
Block Matching Preprocessing. Sensors, 21(11), 3855.

[16] Kaur, G., & Singh, M. (2022). "A Comparative Evaluation of CNN Architectures for
Deepfake Detection." Multimedia Systems, 28(5), 703–716.