SemDiffShield: Agentic AI-Empowered Secure Image Semantic Communication via Cryptographic Diffusion

In the emerging era of agentic-AI-driven 6G networks, the exchange of high-dimensional semantic features is a cornerstone of collaborative perception. However, the open nature of wireless channels and the transparency of semantic features create critical vulnerabilities exploitable by model inversion attacks. Existing defenses often rely on rigid bit-level encryption, failing to balance the high-fidelity recovery required by AI inference with robustness against wireless impairments. To address this challenge, we propose SemDiffShield, a framework that repurposes channel noise into an endogenous security primitive. Through three cooperative agents, the system shifts from static encryption to generative trajectory matching. The Diffusion Dynamics Agent enforces a private sampling trajectory that acts as a temporal lock accessible only to synchronized peers. To further remove residual correlations, the Latent Noise Obfuscation Agent applies key-controlled isometric rotations, masking semantic structure while preserving the signal energy essential for channel robustness. Finally, the Semantic Generation Agent ensures transmission undetectability by using large language models to synthesize visually natural cover images semantically orthogonal to the secret source. Empirical results show that SemDiffShield balances confidentiality and resilience, delivering high-fidelity recovery even under severe noise conditions where traditional baselines fail. The source code is available at https://github.com/Daydream0819/SemDiffShield.git