Authors: Shaun Paul Moses, Vignesh. S
Abstract: Another emerging danger in the world of cybersecurity is the term steganography, which means concealing secret data in digital form, because concealed messages can be easily transferred to a different information exchange format. Other modalities such as audio steganography possess unique features that make it difficult to detect such signals, such as the temporal-frequency properties and audio signals are high dimensional. This project offers a DLDA, Deep Learning Based Detection System Stegware in Audio Files, that will inform whether the audio sample is a real cover or it is a stegware, i.e. it has embedded data in it. The system employs improved methods of feature extraction like Spectrogram Analysis and Mel-Frequency Cepstral Coefficients (MFCCs) to identify requisite frequency, amplitude and temporal indications to identify stegmodifications. The CNNs and LSTMs process subsequently learn a discriminative feature (CNNs) and temporal patterns (LSTM) that occurs between normal and manipulated audio. Training and testing are done using a dataset of clean audiofiles and audiofiles with various modifications done using steganography. The performance is measured by the accuracy, precision, recall and F1-score and the system has been found to be very reliable with accuracy of 97.8 and very few false detections. In the experimental results, it is seen that the model works fairly well when noise and compression is introduced, indicating its strength in the real world. Overall, the framework that is created due to the research effectively applies deep learning to offer a scalable, automated and accurate method of audio steganalysis, which is an outstanding achievement that can provide cybersecurity, digital forensics and secure communications as the number of illegal data transmission via audio channels decreases.
