The   growing   sophistication   of   deepfake   technologies   has   emerged   as   acritical threat to the credibility of digital media by generating highly realisticyet  fabricated  visual  content.     This  erodes  public  trust,   elevates  securityvulnerabilities,  and  challenges  information  integrity  across  online  platforms.Despite notable advancements, existing research still suffers from limited datadiversity,  insufficient  model  explainability,  and  inadequate  model  evaluation.To   overcome   this   limitation,   a   framework   for   detecting   deepfake   videomanipulation  by  using  a  transfer  learning  approach  was  introduced.    Eachextracted frame was processed by a convolutional neural network (CNN)-basedmodel to obtain frame-level predictions,  which were subsequently aggregatedto  produce  the  final  video-level  prediction  using  a  predefined  threshold.The  publicly  available,  widely  adopted  FaceForensics++  dataset  was  used,which  contains  high-quality  videos  generated  using  advanced  manipulationtechniques.    Various  CNN  architectures,  including  Xception,  Densenet121,InceptionResNetV2, ResNet50, and EfficientNetB3, were explored along withrigorous  hyperparameter  tuning.     Among  these,  the  Xception  architectureoutperformed others by achieving a test accuracy of 94.5%.  Gradient-weightedclass   activation   mapping   (Grad-CAM),   generalized   gradient-based   visualexplanations (Grad-CAM++), and Shapley additive explanations (SHAP) wereemployed to enhance model explainability by visualizing the key regions thatinfluence  deepfake  detection.    The  research  offers  an  effective  approach  toaddress deepfake threats and safeguard information integrity in contemporaryindustry 4.0