ARCHIVES

There is a lot of concern with how fast artificial intelligence (AI), machine learning, and other technologies have allowed the production of fake, but very realistic synthetic media (deepfakes). Deepfakes create problems with trustworthiness of media, individuals’ privacy rights, and national security. Generative models are rapidly advancing, and especially diffusion based models, are allowing for less noticeable artifacting in manipulated photos; CNNs may not be able to detect these types of photo manipulations as effectively as they used to. In addition to providing a structured review of image based methods for detecting deepfakes using Vision Transformer Architectures (which use self-attention to capture semantic relationship globally across the entire image); we will also provide experimental evaluation of an image-based Vision Transformer architecture for detecting deepfakes generated by current generative models. Experimental results on well established benchmarks and diffusion generated images indicate the accuracy of our approach ranges between 80 – 85%, showing the ability of transformer based models to detect global inconsistency in deepfakes. We will also discuss some challenges to detecting deepfakes including data quality, generalizing to new forms of manipulation, adversarial robustness, and ethics of deepfakes. Additionally, we highlight emerging areas of research, specifically Explainable Artificial Intelligence (XAI), to support development of completely transparent deepfake detection systems. Ultimately, this work highlights the need for Vision Transformer Architecture based approaches to develop robust and future ready deepfake detection systems.