An end-to-end convolutional neural network for secure image transmission via joint encryption and steganography

Area of Science:

• Computer Science
• Information Security
• Digital Image Processing

Background:

• Conventional methods for secure image transmission (encryption, steganography) have limitations, including conspicuous artifacts or vulnerability to attacks.
• Cryptographic algorithms (AES, DES) offer security but produce noticeable image noise.
• Steganographic schemes (LSB, DCT) are imperceptible but lack robust resistance to extraction attacks.

Purpose of the Study:

• To develop an integrated end-to-end convolutional neural network for secure image transmission.
• To fuse image transformation and data hiding into a single, unified network architecture.
• To overcome the drawbacks of conventional sequential pipelines in image-based communication security.

Main Methods:

• An all-new end-to-end convolutional neural network architecture.
• A learnable Key Mixer for adaptive, content-sensitive transformations.
• An Encoder for smooth embedding and a symmetric Decoder for accurate secret recovery.
• A dual-loss operation balancing perceptual fidelity and recovery quality.

Main Results:

• The proposed model achieves Peak Signal-to-Noise Ratio (PSNR) values > 32 dB and Structural Similarity Index (SSIM) values > 0.90.
• Experimental results on STL-10 data demonstrate superior performance over traditional methods in both quality and recovery fidelity.
• Entropy analysis indicates near-randomness, attesting to the security of the embedded data.

Conclusions:

• The developed framework offers a viable and scalable architecture for secure image transmission.
• It establishes a new paradigm by combining content protection with visual invisibility.
• The model effectively balances security, fidelity, and imperceptibility in image-based communication.