Sumit Shekhar Reviewer
Approved
Relevance and Originality
The manuscript addresses a timely problem in AI generated image detection with relevance to digital forensics, cybersecurity, and content authentication. The comparison of EfficientNet B0, EfficientNet B3, and ResNet152V2 is useful. However, the authors should clarify the specific novelty of the study and explain how it advances beyond existing comparative research.
Methodology
The methodology is logically organized and uses consistent training conditions across the three architectures. However, important details about dataset size, image sources, generative models, augmentation, fine tuning, hardware, and software are insufficiently reported. These details should be added to improve reproducibility and demonstrate that the dataset represents diverse synthetic imagery.
Validity and Reliability
The reported accuracy, precision, recall, F1 score, ROC AUC, and confusion matrices provide useful evidence of performance. Nevertheless, the high results require stronger validation. The authors should examine possible overlap between related images or generators across data partitions and consider repeated experiments, confidence intervals, or an independent test set. Testing against unseen generators and common image transformations would provide stronger evidence of generalization.
Clarity and Structure
The manuscript follows a clear research structure and presents the experimental workflow effectively. Some discussion of EfficientNet B3 is repeated across the results, discussion, and conclusion. These repetitions should be reduced. Tables, confusion matrices, model names, and numerical results should be formatted consistently, with clearer captions and labels. Language editing would also improve technical clarity.
Results and Analysis
EfficientNet B3 provides the strongest reported performance, achieving 97.63 percent accuracy and 0.993 ROC AUC. The results are promising, but the analysis should go beyond ranking the models. Parameter counts, training time, inference time, memory usage, or other computational measures should be reported when discussing efficiency. The authors should also analyze misclassified images and compare their findings with recent approaches under comparable conditions.
Overall assessment
The manuscript addresses an important topic and presents promising results. The main revisions needed concern dataset transparency, reproducibility, external validation, generalization, computational evaluation, and clearer positioning of the research contribution.
Sumit Shekhar Reviewer