A provenance-aware hybrid deepfake detection framework integrating perceptual hashing and explainable CNNs for robust multimedia forensics in social IoT environments

The rapid advancement in generative AI technology have increased the problem of detecting deepfakes, especially in the decentralized social Internet of Things (Social IoT), where multimedia data gets modified constantly in various ways such as by compressing, resizing, and metadata removal. Conventional techniques usually attempt to solve either manipulation detection or provenance verification independently, lacking flexibility and efficient implementations capable of running on edge devices. This paper introduces a provenance-aware hybrid deepfake detection framework based on CNN and perceptual hashing (pHash) as an efficient approach for provenance validation. EfficientNet-B0 and ResNet50 are utilized as CNN-based detectors, while an adaptive and confidence-aware fusing technique helps determine the ratio between the confidence of detectors and provenances. Without valid provenance information, the hybrid technique switches to detector-based inference. Grad-CAM was used to visualize the reasoning of the proposed framework. Experiments were conducted on the Deepfake and Real Images dataset using different combinations with fixed train-validation-test splits over three random seeds. The proposed framework achieved a mean AUC of 0.9973 with EfficientNet-B0 under the hybrid setting, while maintaining highly competitive classification performance with ResNet50 across multiple experimental configurations. Cross-validation experiments on the Celeb-DF benchmark dataset further demonstrated strong generalization capability and robustness across multiple random seeds. Provenance survivability evaluation in the case of Social IoT transformations, like resizing, compression, and Gaussian blur, resulted in an average survival probability of 47.05%. It is crucial to note that adaptive hybrid inference should be applied for such tasks. The ablation study, statistical test of significance, computational complexity evaluation, and qualitative Grad-CAM visualization have provided additional proof of the effectiveness and efficiency of the proposed approach. The proposed framework demonstrates strong performance across the evaluated benchmark datasets while maintaining computational efficiency, explainability, and provenance awareness. These findings indicate the potential of the proposed hybrid framework for robust multimedia forensic analysis in decentralized Social IoT environments; however, the results should be interpreted within the scope of the evaluated datasets and dissemination scenarios.