Practical Disruption of Image Translation Deepfake Networks

Nataniel Ruiz, Sarah Adel Bargal, Cihang Xie, Stan SclaroffAAAI 2023 · OpenReview

Abstract

By harnessing the latest advances in deep learning, image-to-image translation architectures have recently achieved impressive capabilities. Unfortunately, the growing representational power of these architectures has prominent unethical uses, including face manipulation ("DeepFakes") and "DeepNude" body image manipulations. While several works tackle the task of disrupting such image translation networks by inserting imperceptible adversarial attacks into the input image, these works often rely on "white-box" knowledge or require an impractical number of queries in "black-box" scenarios. We propose Leaking Transferable Perturbations (LTP), an algorithm that significantly reduces the number of queries needed to disrupt an image translation network by dynamically re-purposing previous disruptions into new query-efficient disruptions.

Introduction

The rise of deepfake technology has created significant challenges for digital privacy and misinformation. Proactive disruption—adding imperceptible noise to an image to "immunize" it against manipulation—offers a promising defense. However, real-world deepfake services typically operate as "black boxes," where the protector has no knowledge of the underlying model's internals. Existing black-box adversarial attacks are often "query-hungry," requiring thousands of interactions with a target API to find a successful perturbation. This makes them slow, expensive, and easy for service providers to detect and block.

Leaking Transferable Perturbations (LTP)

LTP addresses the query efficiency problem by exploiting the observation that adversarial vulnerabilities in image translation models are highly correlated across different inputs. The method operates in two distinct phases:

  1. Leaking Phase: The algorithm first attacks a small "leaking dataset" (e.g., a few dozen images) using standard, computationally intensive black-box methods. This phase collects successful perturbations that disrupt the target model.
  2. Exploitation Phase: Principal Component Analysis (PCA) is performed on the collected perturbations to identify the primary "directions" or patterns of noise that the model is most vulnerable to. When protecting a new image, LTP prioritizes these "leaked" patterns, drastically reducing the search space.
The two-phase LTP process: (1) Leaking Phase to identify model vulnerabilities and (2) Exploitation Phase to efficiently protect new images.

The algorithm can perform two types of disruption: a Neutralizing Attack, which forces the deepfake output to match the original input (effectively nullifying the edit), and a Distortion Attack, which renders the output an unusable mess.

Experimental Evaluation and Results

The researchers evaluated LTP using the CelebA dataset against two prominent architectures: GANimation (expression editing) and StarGAN (attribute editing). LTP demonstrated significant improvements over state-of-the-art black-box attacks such as IT-SimBA and IT-Square.

Query distribution comparison between LTP and IT-SimBA, demonstrating LTP's significantly lower query requirements for successful disruption.

Key results include:

  • Query Reduction: LTP achieved a 3x reduction in queries for facial attribute editing (136 vs. 444 queries) and a 56% reduction for expression editing compared to the next best method.
  • Success Rate: The method maintained a 100% success rate while requiring fewer than 20 queries in many instances, making it nearly instantaneous compared to previous methods.
  • Transferability: The success of LTP confirms that complex adversarial attacks can be decomposed into a few simple linear components that are transferable across different images for a given model architecture.

Conclusion

LTP represents a significant step toward practical image protection against deepfake networks. By slashing the number of required queries, it moves proactive disruption from a theoretical concept to a viable tool for real-world applications, such as automated protection filters for social media uploads. Future work will explore the transferability of these leaked perturbations across different model architectures to create more universal protection mechanisms.

Reference: Ruiz, N., Bargal, S. A., Xie, C., & Sclaroff, S. (2023). Practical Disruption of Image Translation Deepfake Networks. Proceedings of the AAAI Conference on Artificial Intelligence.

Summary updated

Abstract

By harnessing the latest advances in deep learning, image-to-image translation architectures have recently achieved impressive capabilities. Unfortunately, the growing representational power of these architectures has prominent unethical uses. Among these, the threats of (1) face manipulation ("DeepFakes") used for misinformation or pornographic use (2) "DeepNude" manipulations of body images to remove clothes from individuals, etc. Several works tackle the task of disrupting such image translation networks by inserting imperceptible adversarial attacks into the input image. Nevertheless, these works have limitations that may result in disruptions that are not practical in the real world. Specifically, most works generate disruptions in a white-box scenario, assuming perfect knowledge about the image translation network. The few remaining works that assume a black-box scenario require a large number of queries to successfully disrupt the adversary's image translation network. In this work we propose Leaking Transferable Perturbations (LTP), an algorithm that significantly reduces the number of queries needed to disrupt an image translation network by dynamically re-purposing previous disruptions into new query efficient disruptions.
BibTeX
@inproceedings{ruiz2023practical,
  author = {Nataniel Ruiz and Sarah Adel Bargal and Cihang Xie and Stan Sclaroff},
  title = {{Practical Disruption of Image Translation Deepfake Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=c3GLhf18P8z},
}
Figure