AI Based Secure Text and Audio Communication Using Image Steganography

Secure messaging must protect the message content while also controlling how visible the communication is. This paper describes an artifact-validated study of a web prototype that encrypts either text or binary audio using AES-256-GCM and encloses the resulting JSON envelope within a lossless RGB image. The steganography engine compares image-sample complexity using a map constructed from Sobel gradients and local variance, and replaces one LSB in each selected colour-channel sample. A 12-byte header contains a magic number, payload length, and CRC32 checksum. The supplied source also contains Django REST models and JWT authentication code, a Next.js client, a standalone browser client, ECDH utilities, neural-network definitions, and an Isolation Forest training script. However, code tracing reveals that ECDH is not connected to the message path, AES keys are embedded in the envelope, the packaged CNN checkpoints are missing, CNN inference responses are stubs, and runtime threat scoring is heuristic rather than Isolation Forest inference. Tests of the actual adaptive-LSB class retrieved an 81-byte text envelope and an 8,044-byte WAV payload without error from a deterministic 512 × 512 RGB cover; a modified embedded bit caused a CRC mismatch, and a clean image was rejected. The contribution is therefore a clear description of the implemented prototype, its proven functionality, and the modifications necessary for it to offer end-to-end secure communication.