This Face Does Not Exist... But It Might Be Yours! Identity Leakage in Generative Models

Patrick Tinsley, Adam Czajka, Patrick FlynnWACV 2021

TL;DR

This paper investigates whether synthetic faces generated by StyleGAN2 inadvertently leak identity information from their training data. The authors demonstrate that this "identity leakage" does occur for some face recognition models, suggesting that generated faces might not be as anonymous as previously assumed and posing potential privacy risks.

What Problem Does This Paper Address?

Generative Adversarial Networks (GANs), particularly models like StyleGAN2, have become remarkably adept at creating photorealistic synthetic images, including human faces, that appear novel and "do not exist." This capability has exciting applications but also raises concerns about privacy. Specifically, this paper tackles the question of whether the training process of these GANs, which involves learning the distribution of real faces, might inadvertently embed identifiable information from the training set into the generated "fake" faces. If so, this could mean that synthetic faces, while not direct copies, might still carry traces of real individuals, potentially leading to privacy breaches or misidentification when used with face recognition systems.

Methods

The core of the study involves training a StyleGAN2 model on the FFHQ dataset and then evaluating the degree of identity leakage using five different face recognition models: face_recognition, FaceNet, ArcFace, SphereFace, and Neurotechnology MegaMatcher. The methodology centers on comparing "Real-Real" (R-R) face image pairs (both images from the original training set) with "Real-Generated" (R-G) face image pairs (one real image from the training set, one synthetic image from the StyleGAN2 model).

The researchers analyze the comparison scores produced by each face matcher for these pairs. If identity leakage is present, the R-G pairs should, on average, be considered more similar (i.e., have lower distance scores for most matchers) than R-R pairs. This is because the synthetic image would have inherited identity-salient features from the real training image. To further investigate, they also analyze the "closest image" grids, which visually show which synthetic faces a matcher deems most similar to a given real face.

Datasets

  • FFHQ (Flickr-Faces-HQ): A dataset of 70,000 high-resolution (1024x1024) images of human faces, used by NVIDIA to train StyleGAN2 and the original StyleGAN. Each image represents a different individual.
  • SG2 (StyleGAN2-generated fake face images): A newly generated dataset of 70,000 synthetic face images created using StyleGAN2 with a truncation value of 0.5. This dataset is made publicly available for replication.
  • UNI: A dataset of 12,004 high-quality face images collected by the authors, featuring 333 subjects with multiple front-facing photos. This dataset was used for preliminary performance analysis of the face matchers and is subject-disjoint from FFHQ and SG2.

Key Insights and Ideas

  • Identity leakage is real, but not universal: The paper definitively shows that identity information can leak from training data into synthetic faces generated by StyleGAN2. However, this leakage is not consistent across all face recognition models.
  • Matcher architecture matters: The degree of identity leakage appears to be strongly dependent on the specific face recognition model used. Some models are highly susceptible, while others show little to no evidence of leakage.
  • Privacy implications are significant: The observed leakage, even if subtle, suggests that synthetic faces might not be truly anonymous. This has direct implications for privacy, especially in scenarios where synthetic faces are used to create fake identities or impersonate individuals.
  • "Closest image" analysis provides qualitative insights: Visualizing the synthetic images that face matchers consider "closest" to real faces helps explain why certain matchers detect leakage, revealing how they interpret facial similarities.

Key Results

  • ArcFace and SphereFace show strong evidence of identity leakage: For these two models, the "Real-Generated" (R-G) comparison scores were significantly shifted towards higher similarity (smaller distances) compared to "Real-Real" (R-R) scores. This indicates that synthetic faces generated by StyleGAN2 are perceived as more similar to real faces from the training set by these matchers.
  • face_recognition and FaceNet show minimal to no leakage: These models exhibited strong overlap between R-R and R-G score distributions, suggesting they are robust against this specific type of identity leakage.
  • MegaMatcher 11.2 also shows minimal leakage: Despite being a high-performing commercial matcher, it did not show significant evidence of identity leakage in this context.
  • Privacy threat quantified: Using ArcFace as an example, the paper illustrates that a real user whose face was in the training data would be approximately 6.6 times more likely to be falsely matched than someone not in the training data, highlighting a practical privacy threat.

Limitations and Future Work

  • Inconclusive reasons for variation: The paper acknowledges that the exact reasons why identity leakage varies so much between face matchers remain unclear. This is a significant area for future research.
  • Limited scope of StyleGAN2: The study focused on StyleGAN2. Future work could explore other generative models like StarGAN, CycleGAN, and GauGAN.
  • Further investigation of matcher geometry: The authors suggest that the linearity vs. non-linearity of face matcher geometry might play a role in identity leakage, warranting further rigorous exploration.
  • Quantitative analysis of feature extraction: Using tools like GradCAM to visualize and compare feature extraction between leakage-prone and leakage-resistant matchers is proposed for deeper understanding.
  • Testing differential privacy solutions: The paper suggests validating differential privacy-based GANs on larger datasets to assess their effectiveness in preventing unwanted identity leakage.

Summary updated

BibTeX
@inproceedings{tinsley2021this,
  author = {Patrick Tinsley and Adam Czajka and Patrick Flynn},
  title = {{This Face Does Not Exist... But It Might Be Yours! Identity Leakage in Generative Models}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2021},
}
Figure