Abstract
Semantic Communication (SemCom) is a transformative paradigm for 6G networks, prioritizing the transmission of "meaning" over raw bit-level accuracy to achieve unprecedented spectral efficiency. However, this efficiency introduces a significant security paradox: the highly compressed semantic features transmitted across the channel are inherently descriptive of the source data. This paper investigates the Model Inversion Eavesdropping Attack (MIEA), demonstrating how an adversary can reconstruct private source images from intercepted semantic features. To mitigate this threat, we propose a lightweight defense mechanism based on random permutation and substitution that secures data without the prohibitive overhead of traditional cryptographic methods.
Introduction
Traditional communication systems focus on the technical problem of accurately delivering bits. In contrast, Semantic Communication leverages Deep Learning (DL) to extract essential features from a message, discarding redundant information to save bandwidth. While effective for efficiency, these semantic features are "leaky" by nature. Because they represent the core essence of the source data (e.g., facial features or textual context), an eavesdropper with a compatible DL model can potentially reverse the encoding process. This paper identifies and analyzes the vulnerability of SemCom systems to model inversion attacks and provides a robust, low-complexity defense strategy.
The Threat Model: Model Inversion Eavesdropping Attack (MIEA)
The study explores two primary attack vectors based on the eavesdropper's (Eve's) knowledge of the system:
1. White-box Attack
In a white-box scenario, the attacker possesses full knowledge of the transmitter's encoder architecture and parameters. The reconstruction is framed as an optimization problem: Eve seeks to find a "guess" image that, when passed through the known encoder, produces features identical to the intercepted ones. To ensure the reconstructed output resembles a natural image rather than high-frequency noise, the authors employ Total Variation (TV) regularization during the optimization process.
2. Black-box Attack
In a black-box scenario, the attacker has no knowledge of the specific model parameters. Instead, Eve trains a separate "inverse network" using a proxy dataset. This network learns a mapping from the intercepted semantic symbols back to the raw pixel space. This approach demonstrates that even without proprietary model access, the fundamental structure of semantic features remains a liability.
Proposed Defense: Permutation and Substitution
To protect user privacy without sacrificing the performance gains of SemCom, the authors introduce a two-stage obfuscation process:
- Random Permutation: The order of the feature tensors is shuffled based on a secret key shared between the sender (Alice) and the receiver (Bob).
- Substitution: Specific segments of the feature tensor are swapped with features derived from a different image or random noise.
This dual-layer approach ensures that even if an attacker intercepts the signal, the lack of the secret key prevents them from reassembling the features into a coherent structure. While simple shuffling (permutation) often leaves behind a visible "contour" or silhouette of the subject, the addition of substitution effectively destroys the visual information for unauthorized parties.
Experimental Evaluation
Methodology
The research utilized the CelebA dataset, consisting of over 200,000 celebrity face images. Images were processed at a resolution of $180 \times 180$ pixels. This dataset serves as a rigorous benchmark for testing the fidelity of image reconstruction under attack.
Key Results
- Attack Potency: In a standard environment (20dB SNR), the white-box attack achieved a Peak Signal-to-Noise Ratio (PSNR) of approximately 26.3dB, while the black-box attack reached 27.0dB. In both cases, the reconstructed faces were clearly recognizable, confirming a successful privacy breach.
- Defense Performance: The proposed defense significantly degraded the attacker's capabilities. With permutation and substitution active, the eavesdropper's reconstruction quality dropped to a PSNR between 7dB and 13dB, resulting in unrecognizable visual noise.
- System Integrity: Crucially, the intended receiver (Bob) maintained high-quality reconstruction (PSNR > 30dB) by using the secret key to reverse the obfuscation, proving that the defense is both effective and transparent to authorized users.
Discussion and Future Work
The findings highlight that in SemCom, efficiency is often the enemy of privacy. The removal of "redundant" bits—which in traditional systems act as a form of natural noise—makes the transmitted signal a pure, vulnerable representation of private data.
While the proposed defense is effective, future research must address Key Management for large-scale 6G deployments and the Complexity of Substitution when a sender has only a single image to transmit. Furthermore, extending the MIEA framework to other modalities, such as speech and natural language processing, remains a critical next step in securing the future of semantic networking.
Summary updated