KMAN Publication Inc. AI and Tech in Behavioral and Social Sciences In Press 2027 01 01 Compression-Resilient Deepfake Detection for AI-Assisted Social Media Content Monitoring: Feature-Level Fusion of ResNet50 and EfficientNet-B0 with a Learnable Visibility Matrix Compression-Resilient Deepfake Detection for AI-Assisted Social Media Content Monitoring: Feature-Level Fusion of ResNet50 and EfficientNet-B0 with a Learnable Visibility Matrix 1 14 EN Journal Article 2026 04 24 <p>Deepfakes circulating through social media create a combined technical, behavioral, and governance challenge because manipulated videos may be redistributed after resizing, re-encoding, and compression. This study evaluated a frame-level detector combining ResNet50 and EfficientNet-B0 through feature-level fusion and an iterative Visibility Matrix procedure. Frames were obtained from FaceForensics++ c23 and Celeb-DF v2 and resized to 128 × 128 pixels. The fused model produced the highest reported point estimates, including 83% validation accuracy, an AUC-ROC of 0.89, balanced accuracy of 0.81, and approximately 18 ms inference time per image. Reported accuracy was approximately 79% on an additional unseen evaluation and approximately 78–80% under a severe-compression condition, compared with approximately 65–67% for ResNet50. These results are descriptive because the available experiment did not report repeated runs, confidence intervals, formal statistical comparisons, or a fully specified external-testing protocol. The detector should therefore be interpreted as a candidate frame-level triage component requiring video-disjoint validation, formal ablation, calibrated thresholds, and human review before operational use.</p> https://www.journals.kmanpub.com/index.php/aitechbesosci/article/download/5856/11452