Reconstruction-Resistant Image Transmission Using Semantic Communications

Semantic communication has emerged as a promising paradigm for next-generation wireless networks, offering substantial efficiency gains by prioritizing the transmission of task-relevant meaning over bit-level accuracy. However, while its benefits in bandwidth reduction and intelligent data representation are well established, its potential to provide intrinsic reconstruction resistance without relying on conventional cryptographic mechanisms remains largely unexplored. This paper investigates whether semantic communication system architectures themselves can contribute to intrinsic reconstruction resistance for image transmission. We propose an autoencoder-based semantic communication framework in which images are encoded into latent representations and transmitted over a wireless channel, with decoding performed using architecture-specific neural networks. Unlike traditional secure communication approaches that depend on encryption, the proposed method leverages architectural uniqueness and representation-level abstraction to limit unauthorized reconstruction. To systematically analyze this, we evaluate eight adversarial scenarios encompassing variations in encoder–decoder architecture and initialization, including both matched (worst-case) and maximum mismatched (best-case) conditions. The system is modeled using a standard Alice–Bob–Mallory framework, where an adversary attempts to reconstruct intercepted semantic representations without full architectural knowledge. Performance is evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) for reconstruction quality, alongside semantic accuracy measured via a convolutional neural network (CNN)-based classifier and embedding cosine similarity to assess information leakage. Experimental results demonstrate that architectural mismatches substantially degrade both visual reconstruction and semantic interpretability for unauthorized receivers, while matched configurations enable substantial recovery. It is important to emphasise that the proposed approach does not provide cryptographic confidentiality; rather, it offers architecture-dependent resistance to unauthorised semantic reconstruction under restricted adversarial assumptions. Overall, the results show that semantic communication systems can exhibit intrinsic reconstruction resistance through architecture-dependent latent-space organisation, reducing reliance on additional cryptographic overhead under restricted adversarial assumptions, while also highlighting limitations when adversaries possess full architectural and initialisation knowledge.