Deep Fake Detection And Authenticity Verification

Abstract

Deepfake Technology Threatens Information Integrity By Enabling The Creation Of Highly Realistic Synthetic Media That Can Mislead Viewers, Manipulate Public Opinion, And Compromise Biometric Authentication Systems. Existing Detection Models Struggle To Generalize Across Unseen Manipulation Techniques, Suffering From Severe Performance Degradation When Tested On Deepfakes Generated By Methods Not Present In Their Training Data. This Project Leverages The Semantic Understanding Of CLIP (Contrastive LanguageImage Pretraining) With ParameterEfficient FineTuning (PEFT) To Detect Deepfakes More Robustly. Unlike Traditional Deep Learning Approaches That Require Full Model Retraining, Our Method Finetunes Only A Small Fraction Of Parameters (less Than 1%) While Preserving CLIP's Powerful Zeroshot Capabilities. The Framework Processes Face Images Through A Dualencoder Architecture That Compares Visual Features Against Learned Textual Prototypes Of "real" And "fake" Classes. Evaluation Is Performed Across Three Benchmark Datasets—FaceForensics++, CelebDFv2, And DFDC—to Validate Generalization Performance. Experimental Results Demonstrate That Our Approach Achieves Average Crossdataset Accuracy Of 89.7%, Significantly Outperforming Stateoftheart Methods By 8–12% On Unseen Manipulation Types. ParameterEfficient FineTuning Reduces Training Memory Footprint By 85% Compared To Full Finetuning, Making The Solution Practical For Deployment On Standard Hardware. This Work Establishes CLIPPEFT As A Scalable, Generalizable, And Resourceefficient Framework For Deepfake Detection And Authenticity Verification.

Keywords