Enhanced Attribution Algorithms for Deepfake Image Detection

Deepfakes are artificially generated images pose significant risks to media authenticity and security. Traditional detection methods often lack transparency, making it difficult to understand their decision-making processes. This work introduces a novel detection system that enhances interpretability by utilizing advanced attribution techniques. The system employs the InSwapper model to generate Deepfake images, combined with a fine-tuned Residual Networks(ResNet101) for classification. Key attribution methods Integrated Gradients, Guided Integrated Gradients, Adversarial Gradient Integration, and Integrated Decision Gradients are integrated to produce saliency maps that identify the most influential regions in the images affecting Deepfake detection. Developed using Python and PyTorch, this approach significantly improves interpretability. Results indicate that all attribution methods can be successfully employed for detecting Deepfakes, providing clearer insights into the detection process. This advancement paves the way for future developments in real-time detection and broader accessibility in addressing manipulated content.