Detection of sperm whale covert communications using deep learning

Recent advances in covert underwater acoustic communication have demonstrated the feasibility of embedding signals within marine mammal sounds, such as dolphin clicks and whale vocalizations. While most research focuses on developing concealment techniques, limited work addresses the detection of such covert embeddings. This study investigates the detection of covert messages hidden within authentic sperm whale ( Physeter macrocephalus ) vocalizations. Using audio from the Watkins Marine Mammal Sound Database, covert messages were embedded using short-duration, high-frequency chirps masked by natural whale clicks and codas. The chirps corresponded to bits using Baudot code to represent characters. Acoustic features, including high-frequency band energy and spectral variance, were extracted as two-dimensional feature vectors. A Siamese neural network was trained on 690 paired feature samples to classify authentic versus embedded sperm whale audio. The model achieved an accuracy of 97.10% with an F1 score of 96.22%. The results highlight the vulnerabilities of marine acoustic environments and contribute to securing underwater communication environments from adversarial acoustic masking.