Implicit neural representation steganography scheme for secret model transmission

As deep neural network models are increasingly shared over open networks, parameter-level transmission of complete model states through neural carriers has become an emerging problem in model steganography. Existing steganography methods based on implicit neural representations (INRs) mainly focus on multimedia payloads and are not directly designed for transmitting specified neural network model states. To address this problem, we propose an INR-based steganography scheme for secret-model transmission. The proposed method uses standard SIREN networks fitted to public cover images as carriers and segments the floating-point parameter state of a secret model across multiple INR files without modifying the carrier architecture. A key-controlled frozen-slot mechanism preserves the embedded values during cover-image fitting. To accommodate heterogeneous model parameters, we further introduce parameter-type-aware normalization for ordinary weights and normalization statistics, followed by layer-aware distribution mapping according to clean-INR parameter distributions. Experiments across multiple vision-model families, including convolutional networks and a vision transformer, show that, under uncompressed and untampered FP32 transmission, the recovered models match the original task performance at the reported evaluation precision while maintaining high-quality cover reconstruction. Parameter-level recovery fidelity is further quantified using MAE, MaxErr, bit-exact fractions, and parameter-content hashes. Matched-checkpoint comparisons with model-steganography baselines characterize recovery fidelity, transmission cost, and sender-side optimization requirements. Additional experiments characterize the behavior of the proposed scheme under precision conversion, quantization, model-file rewriting, fine-tuning, noise, and pruning.