Hear me if you can! Audio Steganography

Audio steganography is a technique for concealing the existence of information by embedding it within non-secret audio, called the carrier audio signal. There is a trade-off between the amount of information encoded and the imperceptibility of the change in the encoded audio. In this work, we implement a deep-learning based audio steganography technique. An ASR model is trained on the TIMIT dataset. We exploit the susceptibility of the trained model to adversarial examples. Given an input recording and a list of phones to encode, the technique searches for a small imperceptible perturbation. When this perturbation is added to the original audio and passed through the model, we recover the encoded text as the ASR output. PESQ score is used as the evaluation metric to quantify the amount of degradation. We also study the time required for encoding the text as a function of it’s length, nature and PESQ. Lastly, we study the effect of a Gaussian Noise channel on our encoding technique.