Benchmarking Deepfake Detection Robustness Using a Hybrid CNN and Content Provenance Framework

Abstract

Deepfake detection remains a critical task in digital media forensics because synthetic facial images can undermine the reliability of visual evidence and public communication. This study develops and evaluates a hybrid deepfake detection framework that combines convolutional neural network-based visual artefact analysis with content provenance verification. The primary experiments used the Deepfake vs Real 60K dataset, which contained 57,071 facial images divided into real and fake classes. Images were resized to 224 × 224 pixels, normalised and augmented during training. The classification component used an EfficientNetB0-based convolutional neural network with a Squeeze-and-Excitation attention mechanism. The provenance component extracted deep feature embeddings and compared input images with real and fake reference databases using cosine similarity. The CNN probability and provenance probability were combined using weighted score fusion to generate the final prediction. The CNN with attention achieved 93.5% accuracy and a ROC-AUC of 0.9937. The provenance module achieved 98.8% accuracy and a ROC-AUC of 0.9996 under controlled within-distribution evaluation. The hybrid framework achieved 94.3% accuracy, precision of 0.95, recall of 0.94, F1-score of 0.94 and ROC-AUC of 0.9969. Provenance score analysis showed clearer separation between real and fake images than perceptual hashing. Cross-dataset evaluation indicated variable generalisation performance, with accuracy ranging from 52.30% to 99.32% across external datasets. The findings suggest that provenance-based similarity analysis can complement CNN-based detection and improve interpretability. However, performance depends on dataset distribution and reference database coverage, and further validation with independent datasets is required before operational use in digital forensic workflows.