Enhancing Copy–Move Forgery Detection with Contrastive Pretraining and Tasmanian Devil-AdamW Optimization

Abstract

‌Copy–move forgery detection is challenging due to deep, subtle models, limited annotated data, and hyperparameter tuning. We present a robust framework combining contrastive self-supervised pretraining with dynamic multi-step optimization to ensure stable convergence, high accuracy, and improved generalization. In the first step, a backbone network is pretrained by applying a contrastive paradigm (BYOL, SimCLR, MoCo v3) on broadly tagged video frames augmented with forgery-aware transformations, including mild geometric changes, compression, synthetic copy–move, and blurring. This step learns invariant and discriminative visual representations without heavy dependence on tags. In the second step, the pretrained encoder is fine-tuned with a lightweight temporal head to capture the spatio-temporal inconsistencies indicative of copy–move manipulations. Training is performed using a dynamic hybrid optimizer: The Tasmanian Devil Optimizer (TDO) performs early-step global exploration across mild architectural knobs and hyperparameters, after which optimization switches to AdamW to guarantee fine convergence and effective exploitation, mitigating local minima. Broad assessments (such as GRIP and VTD) show stable improvements across robust baselines in AUC, accuracy, and F1, while attention maps provide interpretable localization of tampered areas. The proposed model reduces reliance on labels, enhances robustness, and exhibits faster and more stable training dynamics.