Indonesian J our nal of Electrical Engineering and Computer Science V ol. 37, No. 2, February 2025, pp. 801 ∼ 810 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v37.i2.pp801-810 ❒ 801 A deep lear ning model with an inducti v e transfer lear ning f or f or gery image detection Prabhu Be vinamarad 1 , Prakash H. Unki 2 , V enkatesh Bhandage 3 1 Department of Computer Science and Engineering, BLDEA ’ s V .P . Dr . P .G. Halakatti Colle ge of Engineering and T echnology (Af liated to V isv esv eraya T echnological Uni v ersityBelag a vi), V ijayapur , India 2 Department of Information Science and Engineering, BLDEA ’ s V .P . Dr . P .G. Halakatti Colle ge of Engineering and T echnology (Af liated to V isv esv eraya T echnological Uni v ersityBelag a vi), V ijayapur , India 3 Department of Computer Science and Engineering, Manipal Institute of T echnology , Manipal Academy of Higher Education, Manipal 576104, Karnataka, India Article Inf o Article history: Recei v ed Apr 16, 2024 Re vised Sep 25, 2024 Accepted Sep 30, 2024 K eyw ords: Cop y-mo v e for gery Deep learning Image tampering Image-splicing Inducti v e transfer learning ABSTRA CT Due to the a v ailabil ity of af fordable electronic de vices and se v eral adv anced on- line and of ine multimedia content editing applications, the frequenc y of image manipulation has increased. In addition, the manipulated images are presented as e vidence in courtrooms, circulated on social media and uploaded upon au- thentication to decei v e the situation. This study implements a deep learning (DL) frame w ork with inducti v e transfer learning (ITL) by using a pre-trained netw ork to benet from the disco v ered feature maps rather than starti ng from scratch and ne-tuning the process to check and classify whether the suspected image is authenticated or for ged ef fecti v ely . T o e xperiment with the proposed model, we used both Columbian uncompressed image splicing detection (CUISD) and the CoMoF oD dat aset for training and testing. W e measured the model’ s per - formance by changing h yperparameters and conrmed the better selection of v alues for the h yperparameter to yield compromised results. As per the e v alu- ation results, our model sho wed impro v ed results by classifying ne w instances of images with an a v erage precision of 89.00%, recall of 86.43%, F1-score of 87.32, and accurac y of 87.72% and consistently performed better compared to other methods currently in use. This is an open access article under the CC BY -SA license . Corresponding A uthor: V enkatesh Bhandage Department of Computer Science and Engineering, Manipal Institute of T echnology Manipal Academy of Higher Education Manipal 576104, Karnataka, India Email: v enkatesh.bhandage@manipal.edu 1. INTR ODUCTION T oday , multimedia content has become a signicant part of v arious automated digital systems, such as user authentication, and online user document v erication. M ultimedia denes digital content that includes audio, video, images, and combinations. Combining these automated digital syste ms and v arious multimedia data has created state-of-the-art technologies that pro vide v arious digital services to humankind to replace tra- ditional methods and increase producti vity . On the other hand, the incre ased sophistication of image editing tools, for gery applications, and articial intelligence (AI) algorithms caused a swift upsur ge in digitally ma- nipulated counterfeits in internet crime w orldwide [1]. F or instance, criminals carry out criminal acti vities and pose a signicant threat to internet users by tampering wi th images and videos. The y create altered images J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
802 ❒ ISSN: 2502-4752 to misrepresent their meaning with malicious intentions and f ak e e vidence or documents to present in court- rooms without signicantly lea ving visual clues to be detected by the nak ed e ye. There are man y dif ferent w ays to manipulate multimedia content, and such instances o f digitally f abricated photos from the past ha v e been preserv ed in F arid’ s collection of web pages [2]. Think of cop y-mo v e image for gery as one of the for gery e xamples; in this case, some copied parts are placed at another location within the same image. The tw o im- age sources are combined to form a for gery image in image splicing. Similar tampering is also performed in the case of audio and video data [3]–[5]. In all these types of multimedia tampering, image tampering has g ained much attention due to the a v ailability of cameras with e xcellent resolution, state-of-the-art photo-editing applications, and adv anced softw are tools. The rst block-based for gery detection approach w as introduced in [6]. Since then, v arious image tampering detecti on techniques ha v e been de v eloped, incorporating di v erse feature e xtraction and matching methods. W ith the increasing popularity of deep neural netw ork frame w orks, there ha v e been ef forts to utilize con v olutional neural netw orks (CNNs) to enhance for gery image detection. This section summarizes the latest image tampering detection approaches in recent years. Bayar and Stamm [7] introduced a CNN with layers dedicated to suppressing the content of an im- age and learning manipul ation detection features adapti v ely . Ouyang et al. [8] emplo ys a pre-trained model already created from an e xt ensi v e database, such as ImageNet. Then it modies the net structure slightly us- ing tin y training cop y-mo v e samples to nd cop y-mo v e image samples generated automatically by computer . Huang et al. [9] describes a CNN that can understand features e xtracted from each con v olutional layer and autonomously learn features to recognize dif ferent types of im age manipulation. Muzaf fer and Ulutas [10] discusses the detection and localization of cop y-mo v e for geries using a DL-based frame w ork instead of con- v entional feature e xtraction methods. Abdalla et al. [11] suggested a CNN model with added pre-processing layers to detect dif ferent cop y-mo v e for gery images. The e xperiments demonstrate that the total v alidation ac- curac y stands at 90%. Elaskily et al. [12], mention CNN trains hierarchical features represented from an input image to identify altered and original images. Rodriguez-Orte g a et al. [13], proposed tw o approaches, a model using a custom architecture and a model using transfer learning to distinguish between altered and original images. Abbas et al. [14] proposed tw o lightweight SmallerV GGNet (inspired by V GGNet) and MobileNetV2 deep learning (DL) models to classify and detect cop y-mo v e for gery and post-for gery images. Abhishek and Jindal [15] proposes a method based on color illumination, deep CNN, and semantic se gmentation to detect and localize image for geries. Goel et al. [16] introduces a DL approach with a dual-branch CNN to detect passi v e cop y-mo v e for gery by e xtracting multi-scale features using v arious k ernel sizes. Kadam et al. [17] presents a lightweight model constructed using mask re gional CNN (R-CNN) with MobileNet V1 to detect the for geries present in an image along with corresponding percentages. F ahn and W u [18], proposed a DL-based method for detecting for gery images. This method utilizes discrete fourier transform and contrasti v e learning, enabling the model to directly learn the dif ferences between authentic and for ged images. Mehrjardi et al. [19] presents a DL method for image-le v el for gery detection by emplo ying a pre-trained deep model and global a v erage pooling (GAP). Additionally , pix el-le v el for gery detection is achie v ed through heatmap acti v ation. Similar kinds of DL models with transfer learning and particle sw arm optimization(PSO) techniques are discussed in [20]–[22], respecti v ely . Sadanand et al. [23] proposed CNN with error le v el analysis error le v el analysis (ELA) adopted to detect and accurately classify the cop y-mo v e for gery images. He et al. [24], proposed a method for detecting GAN-generated for gery images by combining central dif fere n c e con v olution and v anilla con v olution (CDC-Mix). This approach considers the depth and width features of neural netw orks and analyzes the impact of attention on netw ork performance. Although man y DL models mentioned abo v e ha v e demonstrated promising results in classifying im- ages as for ged or authentic, the y ha v e some dra wbacks, including the need for impro v ed for gery image de- tection frame w orks, generalization to unseen for geries, and the possibility that training DL models for cop y- mo v e for gery detection will tak e a long time and require po werf u l hardw are. Therefore, this paper presents a DL-based model using inducti v e transfer learning and ne-tuning processes to impro v e the model’ s ability to detect altered images and reduce loss or misclassication. The model is trained usi ng high-quality authentic and for gery images with augmented forms to enhance the learning ability . The proposed w ork also presents a better selection of signicant h yperparameter v alues to enhance the training and testing performance and attain a model generalization. Hence, our model can be inte grated with v arious web applications, go v ernment agencies, and social media platforms as a back end to v erify the le gitimac y of image data and ensure that for ged images are not circulated on social media. Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 2, February 2025: 801–810 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 803 2. PR OPOSED METHOD The primary focus of our proposed methodology is to implement a reliable custom DL model using the pre-trained netw ork and apply the inducti v e transfer learning and ne-tuning process to e xtract rele v ant and discriminati v e image characteristics that are well-structured and acceptable for the netw ork to learn and produce better classication results. Therefore, we used the MobileNet V2 model, de v eloped at Google, and pre-trained on the ImageNet dataset, which consisted of 1.4M images and 1,000 classes. W e adopted inducti v e transfer learning to tak e benet of disco v ered feature maps rather than starting them from scratch. On the other hand, the ne-tuning process enables the higher -order feature representations of the base model to mak e them more pertinent to the particular classication task and ef ciently classify the ne w sample images as for gery or authenticated. T ransf er learning and ne-tuning add a si gnicant step to our proposed frame w ork. In induc- ti v e transfer learning, we frame a ne w fully connected layer on top of the pre-trained model and a classier to use the representations learned by a base netw ork and e xtract signicant features from a ne w set of query image samples. On the other hand, the ne-tuning process unfreezes some of a frozen model base’ s top layers and jointly trains the base model’ s nal layers and ne wly added classier la yer to “ne-tune” the higher -order feature representations of the base model to pre v ent an y modications to the weights of base layers during the backpropag ation phenomenon. Figure 1 depicts the schematic design of the suggested model. ImageNet dataset New classifier head Inductive  T ransfer Learning Forgery Pr edicted   classes Pre-trained weights Classifier Head T ar get dataset Fine-tuning Authentic Convoluation layer Dr opout layer Maxpooling layer ReLu layer Figure 1. Schematic design of our proposed DL model with inducti v e transfer learning 2.1. Description of tar get dataset T o e v aluate the goals of the suggested model, we ha v e considered tw o popularly kno wn and publicly accessed datasets, i.e., CoMoF oD [25] and Columbia uncompressed image splicing detection (CUISD) [26]. The former primarily contains 200 cop y-mo v e for gery images of size 512 × 512 and uses visual appearance- related post-processing attacks. As a result, 10,400 image samples, including original, for gery and ground truth images in the CoMoF oD dataset. The latter includes 363 total image samples in the entire dataset. Out of them, spliced images are 180 and 183 are authentic. The dimension of each image ranges from 757 × 568 to 1152 × 768. T able 1 depicts authentic and for ged image samples from tw o tar geted datasets. T able 1. Authentic and for ged image samples from tar get datasets CUISD dataset [26] CoMoF oD [25] Authentic F or ged F or ged Authentic F or ged F or ged A deep learning model with an inductive tr ansfer learning for ... (Pr abhu Be vinamar ad) Evaluation Warning : The document was created with Spire.PDF for Python.
804 ❒ ISSN: 2502-4752 2.2. Implementation details The proposed methodology uses Python 3.8, K eras and T ensorFlo w as a back end toolkit to incorporate inducti v e transfer learning and ne-tuning process and replace the prior pre-trained FC layer with a ne w FC layer that is framed as a model head and attached to the base of the tar get model. Later , the frame w ork’ s primary layers are frozen to a v oid weight updates during backpropag ation. Subsequently , a dense layer with adequate classication outputs and a s oftMax function is added to the global a v erage pooling layer to create a proper frame w ork for accurately classifying the ndings. Since our model intends to identify only tw o classes (for gery or authentic), tw o outputs and SoftMax function are added to the dense layer . The model can handle input images up to 300 × 300 × 3. During preprocessing, the selected image size is reduced based on the model input designed (300 × 300 pix els) and image pix el v alues scaled between 0 and 1 range. Later , the training dataset size increased articially by using the data augmentation technique to create altered v ersions of the images to e xtract signicant image features from al l perspecti v es to impro v e learning and classication ability . The suggested approach uses K eras’ s DataGenerator class to enhance image data. In this step, e v ery batch of training data is randomly rotated, cropped, and resized. The implementation details in terms of pseudo-code are presented in Algorithm 1. Algorithm 1. Proposed for gery detection methodology Requir e: Images from the dataset: D Ensur e: The input image is F or gery or Authentic 1: function R E S I Z E ( D , [ h, w , d ] ) 2: end function 3: function S P L I T ( D , ratio) 4: end function 5: function N O R M A L I Z E ( imag es, [0 , 1] ) 6: end function 7: Load the tar get domain image dataset. 8: D ← T ar get domain image dataset. 9: Preprocess data from both domains: 10: imag es ← R E S I Z E ( D , [300 × 300 × 3] ) 11: N O R M A L I Z E ( imag es, [0 , 1] ) 12: Perform data augmentation such as random rotations, ips, and shifts. 13: Split D : Di vide 80% for training and 20% for v alidation. 14: [ D T , D V ] ← S P L I T ( D , 0 . 2 ) 15: function M O D E L ( D T , D V ) 16: Load the MobileNet V2 model and modify it for inducti v e transfer learning 17: function L O A D M O B I L E N E T V 2 ( input shape, num cl asses, w eig hts = imag enet ) 18: Set the base model’ s layers in a frozen state. 19: base model .tr a i na bl e ← F al se 20: Add custom top layers for transfer learning 21: x ← base model .output 22: x ← G L O B A L A V E R A G E P O O L I N G 2 D 23: x ← D E N S E ( 128 , activ ation = ′ r el u ′ ) 24: Y pr ed ← D E N S E ( num cl asses, activ ation = ′ sof tmax ′ ) 25: Create transfer learning model 26: model ← M O D E L (inputs= base model .input , outputs= Y pr ed ) 27: end function 28: end function 29: M odel ← M O D E L ( D T , D V ) 30: Compile the modied MobileNetV2 model 31: r esul t ← C O M P I L E T R A I N (Model, optimizer = adam, loss = binary cross entrop y) 32: Ev aluate the model and compute classication metrics. 2.3. Model training and testing The training and testing e xperiments are simulated using 290 and 400 (including original and for gery) images from both datasets. Some authentic images from the CoMoF oD data set are added to the CUISD dataset to increase the number of authentic images during training sessions. The dataset is di vided into 80:20 ratios for training and testing. A predetermined set of initial h yperparameter v alues is k ept constant during initial e xperimentation to assess the training and testing outcome. The parameters include input image size of 300 × 300, initial weight set to ”ImageNet, ” 40 epochs, and batch size=32 with early stopping based on minimum v alidation loss with the patience of 10. The opt imizer is set to ”Adam, ” with a learning rate of 0.000001. Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 2, February 2025: 801–810 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 805 Later , during the simulation, e xperiments are repeated by setting a dif ferent v alue for the epoch, batch size and learning rate to identify the signicant v alue at which the proposed DL model achie v es a better classication result. Figure 2 depicts the training and v alidat ion accurac y and the loss curv es obtai ned during the training and v alidation phase. As per the graphs in Figures 2(a) and 2(b) for training and v alidation o v er 40 epochs. The accurac y of the training and v alidation v aries from 0.78 to 0.92. The training accurac y starts around 0.85. Exhibits signicant uctuations throughout the epochs. The accurac y does not sho w a clear increasing trend and v aries between 0.80 and 0.88. The v alidation accurac y starts high, around 0.92, b ut drops slo wly in the rst fe w epochs to belo w 0.80. After the initial drop, it stabilizes and uctuates around 0.88 to 0.90. It appears rela- ti v ely stable b ut with noticeable uctuations compared to the accurac y of the trai ning. On the other hand, the training and v alidation loss for epochs 0 to 40 v aries between a loss of 0.22 and 0.38. The training loss starts around 0.35 and sho ws a general decreasing trend with some uctuations. The loss decreases to approximately 0.25 by the end of the 40 epochs. The v alidation loss starts high, around 0.38, b ut decreases rapidly in the rst fe w epochs to around 0.30 and continues to decrease gradually with some uctuations and sho ws impro v ement o v erall, reaching approximately 0.22 by the end of the 40 epochs. (a) (b) Figure 2. Accurac y and loss curv e (a) training and v alidation accurac y and (b) training and v alidation loss In summary , the tra ining accurac y sho ws high v ariability and gradually e xhibits an upw ard trend, indicating little inconsistent learning. The v alidation accurac y starts high, drops initially , and stabilizes with uctuations. The uctuations in accurac y suggest potential iss ues with the model’ s stability or o v ertting. The training loss decreases o v er time, indicating that the model is learning. The v alidation loss decreases o v erall, suggesting that training impro v es the model’ s performance on unseen data. 3. MODEL EV ALU A TION AND RESUL T DISCUSSION The e xperimental trials were conducted on an Intel core i7 5 . 1 GH z 64-bit processor with 16 GB RAM and a 4 GB GPU to train and test the model. This section includes a description of e v aluation metrics, e v aluation of the suggested model and a discussion of the e v aluation results. 3.1. Ev aluation metrics T o understand the ef fecti v eness of our model, we ha v e adopted e v al u a tion metrics precision (P), recal l (R), harmonic mean (F1-score), and accurac y (Acc). The (1), (2), (3), and (4) dene the computation of these e v aluation metrics. A deep learning model with an inductive tr ansfer learning for ... (Pr abhu Be vinamar ad) Evaluation Warning : The document was created with Spire.PDF for Python.
806 ❒ ISSN: 2502-4752 P = ( T + ) ( T + + F + ) × 100 (1) R = ( T + ) ( T + + F − ) × 100 (2) F 1 = 2 × P × R ( P + R ) × 100 (3) Acc = ( T + + T − ) ( T + + F − + T + + F − ) × 100 (4) Where, true positi v e ( T + ) relates to ho w man y for gery images are correctly classied as a for gery , f alse positi v e ( F + ) refers to ho w man y authentic images are inaccurately disco v ered as for ged images, true ne g ati v es ( T − ) refers to ho w man y authentic images are accurately determined as authentic, and f alse ne g ati v e ( F − ) refers to incorrectly identifying the count of for gery images as authentic. Later , the adequately trained model is tested with 58 unseen for ged images and 80 unseen authentic images from both datasets. Classication results are presented in the confusion matrix depicted in Figure 3 for CUISD sho wn in Figure 3(a) and CoMoF oD sho wn in Figure 3(b) dataset re specti v ely . Performance metrics computed using TP , TN, FP , and FN yield the follo wing results: for CoMoF oD, precision is 83.87%, recall is 92.86%, F1-score is 88.14%, and accurac y is 87.93%; for the Columbia Uncompressed Image Splicing dataset, precision is 94.12%, recall is 80%, F1-score is 86.49%, and accurac y is 87.5%. The detected results sho wing the actual label, the predicted label, and the image name are detailed in T able 2 for CUISD and CoMoF oD dataset respecti v ely . (a) (b) Figure 3. Confusion matrix for test sets from (a) CUISD and (b) CoMoF oD dataset Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 2, February 2025: 801–810 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 807 3.2. Discussion of r esults o v er tar get dataset The observ ed results are compared with other state-of-the-art reference models to assess ho w wel l our model w orks. T able 3 pro vides a detailed summary of reference model specications and corresponding accuracies, and Figure 4 illustrates the performance compari son in terms of precision sho wn in Figure 4(a), recall sho wn in Figure 4(b), F1-score sho wn in Figure 4(c), and accurac y sho wn in Figure 4(d). As per the results, the dual branch CNN [16] achie v es the highest performance with an F1 score of 94 and a perfect recall, making it ideal for tasks where missi ng an y positi v e instance is unacceptable. Our model, DL wi th ITL stands out for its high precision of 89.00 and competiti v e F1 score of 87.32, indicating its ef fecti v eness in scenarios where precision is as important as recall. MobileNet V1 [17] sho ws the lo west F1 score at 64.20, suggesting it is less ef fecti v e than other models. MobileNet V2 [14] performs better , with an F1 score of 84.40, indicating impro v ements o v er V1. CNN v ariants [11] and [20] sho w v aried performance, with [11] achie ving an F1 score of 88.35 and [20] scoring 82.00, reecting recent impro v ements in CNN design and optimization. The CNN with ELA [23] performs well, with an F1 score of 85.90 and the second-highest precision of 88.10, sho wcasing adv ancements in ELA techniques. The perform ance metric F1 score balances precision and recall. The dual branch CNN [16] achie v ed the highest F1 score of 94, indicating superior o v erall performance. On the other hand, our model, with an F1 score of 87.32, is competiti v e and higher than most other models e xcept [11] and [16] and achie v ed the highest precision of 89.00, indicating fe wer f alse positi v es. Hence, our model demon- strates strong performance across all metrics, e xcelling in precision and competiti v e recall, making it a rob ust choice for image forensic applications. T able 2. V isualization of detection results obtained for CUISD and CoMoF oD dataset CUISD dataset CoMoF oD dataset canong3 k odakdcs330 sub 07 canong3 k odakdcs330 sub 09 003 F CA2 001 F CA2 Predicted: F or ged Predicted: F or ged Predicted: Authentic Predicted: F or ged Actual:F or ged Actual: F or ged Actual:F or ged Actual:F or ged canong3 02 sub 01 canong3 02 sub 02 002 F JC2 002 O BC2 Predicted: Authentic Predicted: F or ged Predicted: F or ged Predicted:Authentic Actual:Authentic Actual: Authentic Actual:F or ged Actual:Authentic T able 3. Comparison of model specication and accurac y with reference models Reference models [11] [17] [14] [16] [20] [23] Ours Model name CNN MobilNet V1 MobilNet V2 Dual branch CNN CNN CNN with ELA DL with ITL Dataset used Composite Composite Composite MICC- F2000 Composite CASIA V1.0, CASIA V2.0 and MICC CoMoF oD and CUISD Input image size 64 × 64 512 × 512 224 × 224 700 × 700 128 × 128 128 × 128 300 × 300 T raining set size 1254 2505 2260 1700 2232 3772 552 T esting set size 537 318 565 150 1488 943 138 No. of epoch 7 with 90 iterations 240 100 100 30 30 40 Ov erall accurac y 90% N A 85.6% 96% 87% 81.1% 87.72% A deep learning model with an inductive tr ansfer learning for ... (Pr abhu Be vinamar ad) Evaluation Warning : The document was created with Spire.PDF for Python.
808 ❒ ISSN: 2502-4752 (a) (b) (c) (d) Figure 4. Comparison of performance metrics: (a) precision, (b) recall, (c) F1-score, and (d) accurac y with reference models 4. CONCLUSION Computer vision-based image authentication has man y cutting-edge applications across v arious elds, including banking, forensics, and medicine. The proposed methodology has tried t o deli v er the system to accurately classify the authentic and for gery images for both CUISD and CoMoF oD datasets. Our model implements a DL concept with inducti v e transfer learning and a ne-tuning mechanism to ef ciently utilize the feature maps from the pre-t rained model and classify for gery images. According to the e v aluation’ s ndings, the suggested model yields an a v erage detection accurac y of around 87.72% and precision of 89.00%, recall of 86.43%, and F1-score of 87.32. The comparati v e study in the results and discussion s ection sho ws that it performs better than other models. Currently , the proposed model can only conrm image authenticity without marking the e xact location of the tampering present in an image. In addition, the proposed DL model requires e xtensi v e input data for the training with an enhanced system consisting of GPU conguration to generate good results. Usually , a system with inadequate resources tak es more time and produces poor classication results. Hence, considering these tw o issues mentioned abo v e, the present w ork w ould be e xtended by training lar ge and di v erse image datasets to increase recall and other performance metrics and highlight the for gery image re gions. Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 2, February 2025: 801–810 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 809 REFERENCES [1] A. Chadha, V . K umar , S. Kash yap, and M. Gupta, “Deepf ak e: an o v ervie w , ” Pr oceedings of second international confer ence on computing , communications, and cyber -security: IC4S 2020 , pp. 557–566, 2021, doi: 10.1007/978-981-16-0733-2 39. [2] R. K umari and H. Gar g, “ An image cop y-mo v e for gery detection based on SURF and fourier -mellin transforms, ” 2023 International Confer ence on Articial Intellig ence and Smart Communication, AISC 2023 , pp. 515–519, 2023, doi: 10.1109/AISC56616.2023.10085429. [3] P . R. Be vinamarad and M. S. Shirldonkar , “ Audio for gery detection techniques: present and past re vie w , ” in 2020 4th International Confe r ence on T r ends in Elec tr onics and Informatics (ICOEI)(48184) , Jun. 2020, pp. 613–618, doi: 10.1109/ICOEI48184.2020.9143014. [4] P . R. Be vinamarad and M. U. Mulla, “Re vie w of techniques for the detection of passi v e video for geries, ” International J ournal of Scientic Resear c h in Computer Science , Engineering and Information T ec hnolo gy , v ol. 2, pp. 199–203, 2017. [5] N. A. Shelk e and S. S. Kasana, “Multiple for gery detection in digital video with V GG-16-based deep neural netw ork and KPCA, ” Multimedia T ools and Applications , v ol. 83, no. 2, pp. 5415–5435, Jan. 2024, doi: 10.1007/s11042-023-15561-0. [6] J. Fridrich, D. Soukal, and J. Lukas, “Detection of cop y-mo v e for gery in digital images, ” in Pr oceedings of Digital F or ensic Resear c h W orkshop , 2003, v ol. 3, no. 2, pp. 652–663. [7] B. Bayar and M. C. Stamm, “ A deep learning approach to uni v ersal image manipulation detection using a ne w con v olutional layer , ” in Pr oceedings of the 4th A CM W orkshop on Information Hiding and Multimedia Security , Jun. 2016, pp. 5–10, doi: 10.1145/2909827.2930786. [8] J. Ouyang, Y . Liu, and M. Liao, “Cop y-mo v e for gery detection based on deep learni ng, ” in 2017 10th International Congr ess on Ima g e and Signal Pr ocessing , BioMedical Engineering and Informatics (CISP-BMEI) , Oct. 2017, pp. 1–5, doi: 10.1109/CISP- BMEI.2017.8301940. [9] N. Huang, J. He, and N. Zhu, “ A no v el method for detecting image for gery based on con v olutional neural netw ork, ” in 2018 17th IEEE International Confer ence On T rust, Security And Privacy In Computing And Communications/ 12th IEEE Interna- tional Confer ence On Big Data Science And Engineering (T rustCom/BigDataSE) , Aug. 2018, pp. 1702–1705, doi: 10.1109/T rust- Com/BigDataSE.2018.00255. [10] G. Muzaf fer and G. Ulutas, “ A ne w deep learning-based method to detection of cop y-mo v e for gery in digital images, ” in 2019 Scientic Meet ing on Electri cal-Electr onics & Biomedical Engineering and Computer Science (EBBT) , Apr . 2019, pp. 1–4, doi: 10.1109/EBBT .2019.8741657. [11] Y . Abdalla, M. Iqbal, and M. Shehata, “Con v olutional neural netw ork for cop y-mo v e for gery detection, ” Symmetry , v ol. 11, no. 10, p. 1280, Oct. 2019, doi: 10.3390/sym11101280. [12] M. A. Elaskily et al. , “ A no v el deep learning frame w ork for cop y-mo v efor gery detection in images, ” Multimedia T ools and Appli- cations , v ol. 79, no. 27–28, pp. 19167–19192, Jul. 2020, doi: 10.1007/s11042-020-08751-7. [13] Y . Rodriguez-Orte g a, D. M. Ballesteros, and D. Renza, “Cop y-mo v e for gery detection (CMFD) using deep learning for image and video forensics, ” J ournal of Ima ging , v ol. 7, no. 3, p. 59, Mar . 2021, doi: 10.3390/jimaging7030059. [14] M. N. Abbas, M. S. Ansari, M. N. Asghar , N. Kanw al, T . O’Neill, and B. Lee, “Lightweight deep learning model for detection of cop y-mo v e image for gery with post-processed attacks, ” in 2021 IEEE 19th W orld Symposium on Applied Mac hine Intellig ence and Informatics (SAMI) , Jan. 2021, pp. 125–130, doi: 10.1109/SAMI50585.2021.9378690. [15] Abhis hek and N. Jindal, “Cop y mo v e and splicing for gery detection using deep con v olution neural netw ork, and semantic se gmen- tation, ” Multimedia T ools and Applications , v ol. 80, no. 3, pp. 3571–3599, Jan. 2021, doi: 10.1007/s11042-020-09816-3. [16] N. Goe l, S. Kaur , and R. Bala, “Dual branch con v olutional neural net w ork for cop y mo v e for gery detection, ” IET Ima g e Pr ocessing , v ol. 15, no. 3, pp. 656–665, Feb . 2021, doi: 10.1049/ipr2.12051. [17] K. D. Kadam, S. Ahirrao, K. K otecha, and S. Sahu, “Detection and localizat ion of multiple image splicing using MobileNet V1, ” IEEE Access , v ol. 9, pp. 162499–162519, 2021, doi: 10.1109/A CCESS.2021.3130342. [18] C.-S . F ahn and T .-C. W u, “ A deep-neural-netw ork-based approach to detecting for gery images generated from v arious generati v e adv ersarial netw orks, ” in 2022 International Confer ence on Mac hine Learning and Cybernetics (ICMLC) , Sep. 2022, pp. 115–123, doi: 10.1109/ICMLC56445.2022.9941295. [19] F . Z. Mehrjardi, A. M. Latif, and M. S. Zarchi, “Cop y-mo v e for gery detection and localization using deep learning, ” International J ournal of P attern Reco gnition and Articial Intellig ence , v ol. 37, no. 09, p. 2352012, Jul. 2023, doi: 10.1142/S0218001423520122. [20] T . Muniappan, N. B. A. W arif, A. Ismail, and N. A. M. Abir , “ An e v aluation of con v olutional neural netw ork (CNN) model for cop y-mo v e and spl icing for gery detection, ” International J ournal of Intellig ent Systems and Applications i n Engineering , v ol. 11, no. 2, pp. 730–740, 2023. [21] A. H. Khalil, A. Z. Ghal w ash, H. A.-G. Elsayed, G. I. Salama, and H. A. Ghal w ash, “Enhancing digital image for gery detection using transfer learning, ” IEEE Access , v ol. 11, pp. 91583–91594, 2023, doi: 10.1109/A CCESS.2023.3307357. [22] ¨ O. Kasim, “Deep learning-based ef cient and rob ust image for gery detection, ” Multimedia T ools and Applications , v ol. 83, no. 21, pp. 59819–59838, Jan. 2024, doi: 10.1007/s11042-023-17946-7. [23] V . S. Sadanand, S. S. Janardhana, S. Purushothaman, S. Hande, and R. Prakash, “Con v olutional neural netw ork-based techniques and error le v el analysis for image tamper detection, ” Indonesian J ournal of Electrical Engineering and Computer Science , v ol. 33, no. 2, pp. 1100–1107, Feb . 2024, doi: 10.11591/ijeecs.v33.i2.pp1100-1107. [24] D. He et al. , “MCDC-Net: multi-scale for gery image detection netw ork based on central dif ference con v olution, ” IET Ima g e Pr o- cessing , v ol. 18, no. 1, pp. 1–12, Jan. 2024, doi: 10.1049/ipr2.12928. [25] D. T ralic, I. Zupancic, S. Gr gic, and M. Gr gic, “CoMoF oD - ne w database for cop y-mo v e for gery detection, ” Pr oceedings Elmar - International Symposium Electr onics in Marine , pp. 49–54, 2013. [26] Y . Hsu and S. Chang, “Detecting image splicing using geometry in v ariants and camera characteristics consistenc y , ” in 2006 IEEE International Confer ence on Multimedia and Expo , Jul. 2006, pp. 549–552, doi: 10.1109/ICME.2006.262447. A deep learning model with an inductive tr ansfer learning for ... (Pr abhu Be vinamar ad) Evaluation Warning : The document was created with Spire.PDF for Python.
810 ❒ ISSN: 2502-4752 BIOGRAPHIES OF A UTHORS Prabhu Be vinamarad completed his B.E. in Information Science and Engineering and M.T ech. Computer Science and Engineering from V isv esv araya T echnological Uni v ersity , Belag a vi, Karnataka, India, in 2008 and 2013, respecti v ely . He is pursuing a Ph.D. in Computer Science and Engineering from V isv esv araya T echnological Uni v ersity , Belag a vi, Karnataka, India. His research interests inc lude deep learning, digital images, and audio forensics. He can be contacted at email: prabhube v@gmail.com. Prakash H. Unki completed his B.E. in Electronics and Communication from Karnataka Uni v ersity , Dhara w ad and M.T e ch. a nd Ph.D. in Computer Science and Information Sciences from V isv esv araya T echnological Uni v ersity , Belag a vi, Karnataka, India. He is w orking as a Professor and Head i n the Department of Informa tion Science and Engineering, B.L.D.E.A ’ s V . P . Dr . P . G. H. Colle ge of Engineering and T echnology , V ijayapur . He has 26 years of teaching e xperience in an engineering colle ge. He published 21 research articles in v arious international journals and confer - ences. Currently , he is guiding tw o doctoral students. His research interests include digital image processing, pattern recognition, machine learning, cloud computing, data science, and big data. He is a member of v arious professional societies. He can be contacted at email: prakashunki@gmail.com. V enkatesh Bhandage currently w orking as an Assistant Professor (Senior Scale) in the Department of Computer Science and Engineering at the Manipal Institute of T echnology , Manipal Academy of Higher Education, Manipal. He completed his Ph.D. from V isv esv araya T echnological Uni v ersity , Belag a vi, in 2020. His research interests include image processing, articial intelligence, medical ima ge processing, and machine learning. He has published 12 papers in international journals and conferences. Currently , he is guiding tw o research scholars. Additionally , he serv es as a re vie wer for se v eral journals, including Springer Nature Computer Science (SNCS), PLOS ONE, Scientic Reports, and Signal, Image, and V ideo Processing. He is a member of IEEE, IE, and ISTE. He can be contacted at email: v enkatesh.bhandage@manipal.edu. Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 2, February 2025: 801–810 Evaluation Warning : The document was created with Spire.PDF for Python.