TL;DR
This paper introduces a novel Semantic Communication (SemCom) framework that integrates differential privacy (DP) to protect sensitive semantic features in images transmitted over insecure channels. By using GAN inversion to extract features and employing neural networks to approximate DP application and removal, the system achieves robust privacy while maintaining high reconstruction quality for legitimate users, even against eavesdroppers with similar channel conditions.
What Problem Does This Paper Address?
Semantic Communication (SemCom) aims to improve transmission efficiency by focusing on the meaning of data rather than its raw form. However, transmitting this semantically rich information over public wireless channels raises significant privacy concerns. Existing methods often struggle to balance strong privacy guarantees with high reconstruction fidelity for legitimate users, or they rely on unrealistic assumptions about eavesdroppers. This paper specifically addresses the challenge of protecting sensitive semantic features within images transmitted over wiretap channels, ensuring that an eavesdropper cannot reconstruct or identify private information, while a legitimate receiver can still accurately reconstruct the original image.
Methods
The proposed framework leverages several key techniques:
- GAN Inversion for Semantic Feature Extraction: A pre-trained Semantic StyleGAN is used to invert an input image (X) into a disentangled latent representation (Z). This representation is composed of "common" latent codes (Zcommon) and "private" latent codes (Zprivate) that correspond to sensitive facial features.
- Approximate Differential Privacy (DP) Mechanism: Instead of directly applying DP noise to Zprivate, which can degrade reconstruction quality, the paper proposes an NN-based DP protection module. This module adds "fake" DP noise that approximates a Laplace distribution, aiming to mimic the effect of DP while allowing for better reconstruction.
- NN-based DP Deprotection Module: At the receiver's end, a corresponding NN-based deprotection module is used to effectively remove the added DP noise from the private latent codes, restoring them to a state suitable for reconstruction. This approach tackles the non-invertibility issue of traditional DP.
- NN-based Encryption: To further enhance security, an NN-based encryptor is employed to encrypt the private latent codes using a shared secret key (K1) between the transmitter and legitimate receiver. This adds an extra layer of protection against eavesdropping.
- System Configurations: The paper considers different system configurations:
- System without Encryptor/Decryptor: Focuses on DP protection and deprotection.
- System with Encryptor/Decryptor (Scenario 1 & 2): Incorporates both DP and encryption, with variations in the eavesdropper's knowledge of protected latent codes.
Datasets
- CelebAMask-HQ dataset: Used for training and evaluating the Semantic StyleGAN model and the proposed SemCom system. The dataset consists of 30,000 RGB face images, with 28,000 used for training and 2,000 for testing. Images are resized to 512x512.
Key Insights and Ideas
- Approximating DP with NNs: The core innovation is using neural networks to approximate the DP process (both adding noise and removing it). This circumvents the inherent degradation caused by traditional DP mechanisms, especially for legitimate users.
- Balancing Privacy and Utility: The system demonstrates a flexible trade-off between privacy and reconstruction quality, controlled by the privacy budget (ε). Higher privacy budgets lead to more obfuscated images for eavesdroppers but can still maintain high fidelity for legitimate users.
- Realistic Eavesdropper Model: The paper assumes the eavesdropper has the same channel conditions as the legitimate user, making the security challenge more realistic and the proposed solution more robust.
- Chaotic vs. Misleading Images: The system can generate either chaotic (unrecognizable) or misleading (but still incorrect) images for the eavesdropper, depending on the privacy budget, offering versatile privacy protection.
Key Results
- High Reconstruction Quality for Legitimate Users: The proposed system achieves reconstruction quality for legitimate users that is very close to the benchmark (Direct Transmission without Protection), with minimal degradation even at low privacy budgets. LPIPS scores for legitimate users are consistently low.
- Effective Privacy Protection for Eavesdroppers: The system significantly degrades the eavesdropper's reconstruction quality. LPIPS scores for eavesdroppers are consistently higher than legitimate users, and FPPSR (Face Privacy Protection Success Rate) approaches 1, indicating that reconstructed images are perceived as different from the original.
- Robustness to Channel Conditions: The system maintains its privacy and reconstruction performance across various channel SNRs.
- Impact of Encryption: The addition of the encryptor/decryptor module further enhances security, particularly in Scenario 1 where the eavesdropper has limited knowledge, leading to a larger gap in LPIPS between the legitimate user and the eavesdropper.
Limitations and Future Work
- Approximate DP Guarantee: The paper acknowledges that the DP guarantee is approximate (ε' > ε) due to the optimization balancing privacy and performance.
- Computational Overhead: While not explicitly detailed, training and inference with NN-based modules and GANs can be computationally intensive.
- Future Work: The paper suggests exploring more advanced DP mechanisms and encryption techniques, as well as extending the framework to other data modalities beyond images. The current work focuses on image transmission, and future research could investigate its applicability to text, audio, or video.
Summary updated