DeepForgeryNet: a hybrid CNN-LSTM and transfer learning framework for robust image forgery and deepfake detection

The increasingly lifelike nature of digitally manipulated images, as well as those generated by AI, presents significant problems for both media authenticity and digital trust. Various detection methods mostly depend on the visual content and thus, might miss the subtle forensic traces. This paper focuses on developing a powerful detection tool that leverages both artifact-level and contextual inconsistencies.

In this work, we introduce DeepForgeryNet, an artifact-savvy deep learning model that combines Error Level Analysis, based preprocessing with a hybrid CNN, LSTM network. The preprocessing step focuses on exposing changes in compression and edges, whereas the CNN, LSTM model hybridly captures the spatial and contextual information. The model is trained end-to-end on publicly available benchmark datasets stratifying the data splits.

The approach taken by the method proposed yielded the following results: accuracy at 95.1%, precision at 94.6%, recall at 94.2%, F1-score at 94.4%, and AUC at 0.98. It surpassed a baseline CNN and transformer models, especially in recall. Stability of generalization with accuracy over 92% was observed in cross-dataset experiments.

Such results serve as evidence that the integration of artifact-aware pre-processing with spatial-contextual feature learning can further enhance the reliability of detection of forgery. Even though there are still problems in the detection of very small manipulations and/or when heavily compressed, this work lays the groundwork for a new generation of digital media verification that is trustworthy.