This paper explores leveraging representations of the data distribution learned by diffusion models to improve a downstream task of deepfake image detection. With the recent upsurge in the popularity of generative AI, it has become increasingly common to encounter disinformation in modalities such as language and text, images and speech. However, a significant portion of disinformative content can solely be attributed to deepfake images. Effective countermeasures in the past have relied upon classifying deepfake images based on spatial irregularities, inconsistencies in high frequency content and fingerprint matching with known residuals from popular deepfake generation models. However, as the technology behind deepfakes continues to advance, there is a growing need for robust detection methods and tools to ensure the integrity of visual information and mitigate the risks associated with the spread of misleading or malicious content. Thus, we investigate using diffusion-generated reconstructions and latent space inversion to enhance deepfake detection, adapting to the changing landscape of visual disinformation. We explore the feasibility of using diffusion generated reconstruction, diffusion generated latent space inversion and high frequency feature extraction for improving the performance of detecting deepfakes.