Deep convolutional neural network (CNN) models have excelled in image classification tasks, notably surpassing other methods in challenges like the ImageNet challenge [11] [12] [13] . This success stems from training these networks on extensive datasets containing millions of images. Notably, pre-trained weights of various models are accessible through platforms like Keras, serving as an open resource for researchers. This availability facilitates transfer learning, enabling the application of knowledge gained from solving one problem to similar tasks, thereby enhancing the efficiency of subsequent endeavors. The following diagram

( Figure 1 ) illustrated the proposed framework of this research.

The proposed model acquire data from a dataset which contains both authentic and forged images. Subsequently these data are preprocessed with modified ELA algorithm. Preprocessing contains image classification, segmentation and denoising. Through preprocessing, two subsets named as training subset and validation subset are generated. These data are transferred to the models (CNN model and FNN model) for training, validation and evaluation. The results obtained through evaluation; the results are transferred to SHAP explainer. Finally, the results generated from SHAP explainer are compared to decide the best model for detecting image forgeries. The process and relevant topics are described in the subsequent phases of this paper.

The process of collecting and preparing data for analysis is referred as Data acquisition [14] [15] . It involves identifying the sources of the data, collecting the data, cleaning and preprocessing the data, and storing the data in a format that is accessible and usable for analysis [16] . To collect data for this research, the CASIA 2.0 Image Tampering Detection Dataset was used [13] .

The methodology entails fine-tuning the model by training the entire network with pre-processed images from the CASIA2.0 dataset [13] , which contains both original and tampered images. Standardization involves resizing all images to 224 × 224 pixels and dividing the dataset into training and test subsets. These subsets categorize images as authentic or forged. Preprocessing includes normalization and evaluating compression-induced errors. CNN architecture, aided by ELA [17] , discerns authentic from forged images, with XAI using SHapley Additive exPlanations (SHAP) [3] [9] [18] to elucidate predictions. The outcomes of training, testing, and prediction processes determine the superior architecture for image forgery detection.

Preprocessing, a crucial step in image processing and computer vision applications like forgery detection, recognition, and classification, involves preparing image data for subsequent processing stages such as feature extraction and analysis, aiming to enhance image quality and extract relevant information. Error Level Analysis (ELA), developed by Neal Krawetz [17] , detects image forgery by analyzing errors introduced during image compression, highlighting regions where manipulation likely occurred. Adjustments made during code implementation of the basic algorithm significantly improved forgery detection. The modified algorithm is as follows Table 1 :

The flow diagram of the modified algorithm is as follows ( Figure 2 ):

The modified algorithm enhances efficiency and robustness by addressing memory leaks through proper closure of temporary images and eliminating unnecessary loops. Pre-processing with modified Error Level Analysis (ELA) prepares images by emphasizing regions with differing compression levels [19] , aiding in the extraction of various image characteristics like texture, edges, and corners. This method enhances the efficiency of tasks such as image classification, segmentation, and denoising.

The major differences between the original algorithm and the modified algorithm are the modified one prevents memory leaks by closing the temporary image and also avoids unnecessary loops. Moreover, the modified ELA algorithm provides more consistency than the traditional one during image handling, which helps maintain a more accurate reference for comparison. Traditional ELA algorithm performs redundant image conversion, which adds more computation time than that of modified ELA algorithm. Thus, the modified implementation becomes more efficient and robust.

CNNs, tailored for image and video analysis, mimic human visual processing and excel at identifying complex patterns [12] [20] . They typically consist of convolutional layers for feature extraction, pooling layers for spatial resolution reduction while retaining vital information, and fully connected layers for merging extracted features to produce final outputs like classification labels or segmentation masks [21] [22] [23] [24] [25] (illustrated in Figure 3 ). Convolutional layers slide small filters over input images to extract features, generating feature maps [21] . Pooling layers reduce computational load and prevent overfitting by decreasing feature map resolution [21] . Fully connected layers integrate extracted features to generate final outputs [21] .

The training process of a CNN is governed by several crucial hyperparameters. These include:

FNN, an artificial intelligence model composed of interconnected neurons, consists of input, hidden, and output layers [26] [27] [28] . Neurons within an FNN compute weighted sums of inputs and apply activation functions like sigmoid or ReLU to generate outputs. During training, FNN weights are optimized to minimize the loss function, evaluating prediction performance [26] . The FNN model comprises of the Input Layer receives data, while the Hidden Layers process and learn complex relationships, and the Output Layer produces final network output, varying in neuron count based on problem type [29] . A simple FNN model architecture is as follows ( Figure 4 ):

FNN hyperparameters regulate the training process, crucial for model performance. Key parameters include:

Training, validation, and evaluation are crucial stages in CNN or FNN model development. Training optimizes model parameters to minimize loss on training data, while validation assesses performance on a held-out dataset to prevent overfitting. Evaluation gauges model performance on a separate dataset for an unbiased estimate of generalization.

To train and evaluate the CNN and FNN model, the CASIA 2.0 dataset is partitioned into two subsets: a training set and a validation set. This division guarantees that the model undergoes training on a specific portion of the data and is then assessed on a separate, unseen portion. The conventional ratio for this partition is frequently established at 80% for training and 20% for validation, ensuring an evaluation of the model’s ability to generalize.

The CNN and FNN models’ training, validation, and evaluation in this study involve several steps. These include Model Definition (outlining the model structure), Model Compilation (configuring training specifics like loss function and optimizer), Model Training (which trains the model on designated data), and Model Evaluation (assessing its performance on validation data). Callbacks used to enhance training include EarlyStopping (monitoring validation accuracy and halting if stagnant), ReduceLROnPlateau (adjusting learning rate with no validation improvement), and LearningRateScheduler (customizing learning rate schedule during training).

As AI permeates various facets of life, understanding its decision-making processes becomes crucial amid increasing complexity. Explainable AI (XAI), also known as Explainable Machine Learning (XML), addresses this need by making AI models more transparent and comprehensible. XAI offers methods to elucidate how AI models reach decisions, allowing humans to interpret their reasoning. Techniques like Feature Importance, Partial Dependence Plots, Local Interpretable Model-Agnostic Explanations (LIME), and SHapley Additive explanations (SHAP) tailor explanations to different AI models and applications, with SHAP calculating feature impact on predictions, aiding in identifying influential features in the model’s predictions.

Figure 5 and Figure 6 illustrates the graphs of accuracy (training and validation), as well as the loss (training and validation), for the CNN and FNN models respectively. The CNN model, when combined with modified ELA algorithm, attains a training accuracy of 99.96% and a validation accuracy of 94.21%. On

the other hand, the FNN model achieves a training accuracy of 99.06% and a validation accuracy of 90.65%. The CNN model with ELA also achieves a validation loss of 1.5101, while the FNN model achieves a lower validation loss of 1.0654. On contrary, FNN model is computationally faster than CNN model. The summary of the findings in Figure 5 and Figure 6 are presented in Table 2 for better assimilation of the findings.

Figure 7 and Figure 8 display training and validation accuracy, along with training and validation loss graphs for the CNN and FNN models, respectively. The

summary of the findings in Figure 7 and Figure 8 are presented in Table 3 for better assimilation of the findings.

With the traditional ELA algorithm, the CNN model achieves a training accuracy of 99.62% and a validation accuracy of 93.30%, with a validation loss of 0.9231. Conversely, the FNN model achieves a training accuracy of 98.50% and a validation accuracy of 89.14%, with a lower validation loss of 0.4411. Despite this, the FNN model exhibits faster computational performance compared to the CNN model.

Deep learning models for image forgery detection are evaluated by analyzing performance metrics like precision, recall, and F1-score, illustrated in Figure 9 .

Table 4 presents additional performance metrics for both models utilizing the modified ELA algorithm and the traditional ELA algorithm.

The confusion matrix, a 2 × 2 table, features diagonal elements denoting true positives (TP) and true negatives (TN) for accurate predictions by the classifier, while off-diagonal elements represent prediction errors like false positives (FP) and false negatives (FN). Figure 10 and Figure 11 depict the confusion matrices for the CNN and FNN models with the modified ELA algorithm, respectively.

Figure 12 and Figure 13 display the confusion matrices for the CNN and FNN models with traditional ELA algorithm, respectively.

The model introduced in this study effectively addressed and surpassed the limitations identified in the previously conducted researches addressed in this paper. The proposed model with modified ELA algorithm and CNN model outperformed other models (FNN model with modified ELA algorithm, CNN model with traditional ELA algorithm, and FNN model with traditional ELA algorithm) across all metrics, including training accuracy, validation accuracy, precision, recall, and F1 score. This suggests its superiority in identifying forged images and minimizing false positives and false negatives.

The proposed model could detect all forms of image forgeries with enhanced accuracy. More so, this model could efficiently incorporated SHAP explainer with CNN model and modified ELA algorithm, which could interpret the predictions more precisely. Computational expenses also mitigated by using GPU-based Kaggle kernel.