Introduction
Privacy-Preserving Generative Adversarial Networks (PP-GANs) have emerged as a promising solution for data sanitization. These tools aim to transform sensitive datasets—such as facial images—into versions that retain utility (e.g., facial expressions) while removing private attributes (e.g., identity). However, the security of these systems often rests on empirical privacy audits: if a secondary "discriminator" network cannot identify the private attribute in the sanitized output, the tool is deemed secure.
This paper demonstrates that such empirical checks are fundamentally flawed. We show that a malicious model provider can design a "subverted" PP-GAN that acts as a Trojan horse. This adversarial model produces outputs that appear perfectly sanitized to both human eyes and automated discriminators, yet contain hidden, recoverable secrets using steganographic techniques.
The Adversarial PP-GAN Framework
The proposed attack, built upon the PPRL-VGAN architecture, integrates secret extraction and embedding directly into the neural network's weights. The process involves three distinct stages:
- Secret Extraction: The GAN's encoder is modified with an additional fully connected layer ($FC\_0$). This layer is trained to capture the "secret"—either a discrete user ID (Scheme 1) or a compressed latent representation of the entire input image (Scheme 2).
- Secret Embedding: The extracted secret is hidden within the frequency domain of the sanitized image. By implementing a Discrete Cosine Transform (DCT) as a differentiable neural layer, the adversary can manipulate the Least Significant Bits (LSBs) of middle-frequency coefficients. This ensures the hidden data is robust and invisible to the human eye.
- Innocuous Architecture: To evade detection by users inspecting the model's structure, the adversarial paths are merged into a standard convolutional architecture, making the subverted model look like a standard multi-layer network.
Experimental Evaluation
The attack was evaluated on two facial expression datasets: FERG (stylized characters) and MUG (real human faces). The results confirm that the subverted models successfully balance three competing goals:
- Utility Preservation: The models maintained 100% accuracy in facial expression recognition, matching benign performance.
- Privacy Check Evasion: The subverted images passed both "weak" and "strong" privacy audits. Discriminators trained to identify users from the sanitized images performed no better than they did on honestly sanitized data (e.g., ~30% accuracy on FERG).
- Secret Recovery: The adversary achieved near-perfect recovery. In Scheme 1, user IDs were recovered with up to 100% accuracy. In Scheme 2, the adversary reconstructed the original private images with a Mean Squared Error (MSE) of $3.36 \times 10^{-4}$, producing visually near-identical replicas.
Conclusion
The success of these subversion attacks highlights a critical weakness in current privacy-preserving machine learning: the reliance on empirical "cat-and-mouse" games. Passing a privacy check does not guarantee the absence of sensitive information; it only indicates the failure of a specific detection method. To build truly secure systems, the community must move toward formal privacy frameworks, such as Differential Privacy, and exercise extreme caution when deploying third-party "black-box" sanitization tools.
Summary updated