A Generalized Analytic for the Detection of Synthetic Media
Abstract
Convolutional neural networks (CNNs) and generative adversarial networks (GANs) have quickly become leading tools for the creation of convincing synthetic images. Such images increase the difficulty of discerning fact from fiction in the information space, where such challenges can degrade the quality and timeliness of decision-making. To compete, we must develop tools that can automatically detect artificially generated images. A major challenge in this area centers around the high number of unique image generation methods. We therefore seek a classification analytic that can successfully generalize when tested on images from multiple image generation algorithms. The 2020 paper "CNN-Generated Images Are Surprisingly Easy to Spot... For Now" by Wang et al. proposes such an approach. The study conducted here independently tests and validates this analytic in a variety of use cases. We begin by focusing on the reproducibility of the analytic using both publicly released and retrained models, the performance of the analytic on a dataset of images where generator type is unknown, and the analytic's effectiveness in the detection of traditional deepfakes. We also examine the analytic's robustness in response to reductions in image quality via compression and adversarial perturbations. Finally, we attempt to improve the analytic's performance by using a state-of-the-art generator to produce a new image training set.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2021
- Accession Number
- AD1151115
Entities
People
- Patrick L. Reilly
Organizations
- Naval Postgraduate School