[Paper Review] Source Generator Attribution via Inversion

Michael Albright, Scott McCloskey|arXiv (Cornell University)|Jan 1, 2019

Digital Media Forensic DetectionComputer Science17 citations

TL;DR

This paper proposes a white-box method for attributing synthetic images to their source Generative Adversarial Network (GAN) by inverting the generation process. By optimizing a latent code that reconstructs the input image, the method simultaneously verifies generator origin and recovers a plausible latent input, enabling precise source attribution for synthetic imagery.

ABSTRACT

With advances in Generative Adversarial Networks (GANs) leading to dramatically-improved synthetic images and video, there is an increased need for algorithms which extend traditional forensics to this new category of imagery. While GANs have been shown to be helpful in a number of computer vision applications, there are other problematic uses such as `deep fakes' which necessitate such forensics. Source camera attribution algorithms using various cues have addressed this need for imagery captured by a camera, but there are fewer options for synthetic imagery. We address the problem of attributing a synthetic image to a specific generator in a white box setting, by inverting the process of generation. This enables us to simultaneously determine whether the generator produced the image and recover an input which produces a close match to the synthetic image.

Motivation & Objective

• To address the growing challenge of identifying the source generator of synthetic images created by GANs, especially in the context of deepfakes.
• To extend traditional image forensics to synthetic imagery, where existing attribution methods are limited.
• To develop a white-box solution that simultaneously verifies generator origin and recovers a latent input matching the synthetic image.

Proposed method

• The method inverts the GAN generation process by optimizing a latent code to reconstruct a given synthetic image.
• It uses a differentiable forward pass through the generator to compute gradients for latent code optimization.
• The optimization minimizes a reconstruction loss between the generated image and the input synthetic image.
• The approach operates in a white-box setting, assuming full access to the generator's architecture and parameters.
• Reconstruction success indicates that the generator could have produced the image, enabling source attribution.
• The recovered latent code serves as a fingerprint for the generator, enabling identification of the source model.

Experimental results

Research questions

• RQ1 Can the inversion of a GAN's generation process reliably attribute a synthetic image to its source generator in a white-box setting?
• RQ2 To what extent can the latent code be recovered such that the generated image matches the input synthetic image?
• RQ3 Does the ability to invert the generation process enable reliable source attribution across different GAN architectures?

Key findings

• The method successfully recovers a latent code that generates a synthetic image closely matching the input, demonstrating high reconstruction fidelity.
• The inversion process confirms whether a given generator could have produced the image, enabling source attribution with high confidence.
• The recovered latent code acts as a unique signature, allowing identification of the specific generator used.
• The approach is effective in a white-box setting, where full access to the generator model is available.
• The method simultaneously verifies generator origin and reconstructs a plausible input, achieving dual objectives in one optimization.