PhD Thesis - Results - Adaptive speech steganography based on multilevel SVD factorization using Hankel matrices and DCT

Dataset Description: Empirical Evaluation of the Adaptive Hankel-based Multi-layer Stegosystem (AHMS) Overview This dataset contains the comprehensive experimental evaluation results for the doctoral dissertation titled "Adaptive speech steganography based on multilevel SVD factorization using Hankel matrices and DCT". It encompasses the raw metric outputs and feature extraction sets used to validate the proposed AHMS architecture against established baseline methods across the steganographic requirements of transparency, robustness, capacity, and security. Experimental Setup The data was generated using the CMU Arctic database, consisting of phonetically balanced, 16 kHz, 16-bit linear PCM English utterances. The evaluation compares five variants of the proposed structured-matrix approach (PC Xue V1, PC Xue V2, PC V1, PC V2, PC V3) against two unstructured reference algorithms (Wen et al. and Xue et al.). The embedding was targeted within the 1.5 kHz to 2.5 kHz frequency band. Directory Structure and File Contents BASE METRICS - PESQ - BER - 50 BPS - ALPHA 0.1 - 1.zip Description: Contains preliminary parameter optimization data. Contents: Evaluates the relationship between embedding strength and signal integrity at a fixed 50 bps capacity. It includes Bit Error Rate (BER) and Perceptual Evaluation of Speech Quality (PESQ) metrics as a function of the scaling factors alpha and beta, sweeping from 0.1 to 1.0. BASE METRICS -- 25 - 100 BPS - 0.7 ALPHA_BETA.zip Description: Contains the core transparency, capacity, and efficiency evaluation metrics. Contents: Includes mean squared error (MSE), peak signal-to-noise ratio (PSNR), PESQ, and short-term objective intelligibility (STOI) data. The metrics are recorded across varying embedding capacities from 25 bps to 100 bps, utilizing the optimized scaling parameters alpha = beta = 0.7. It also contains the computational efficiency logs (embedding and extraction times). ROBUSTNESS - ALL ATTACKS - 50BPS - 0.7 ALPHA_BETA.zip Description: Contains message recovery accuracy data under adverse signal processing conditions. Contents: Logs the BER for all tested methods operating at a 50 bps payload. The attack vectors simulated include MPEG-1 Layer III (MP3) compression (64, 96, 128 kbps), additive white Gaussian noise (AWGN at 30 dB SNR), amplitude scaling (70% and 130%), low-pass and high-pass filtering, and requantization/resampling. FOR CLASSIFIER - 2026-01-03_23-35-01_3Q9VMVA.zip Description: Contains the multidimensional feature spaces extracted for machine learning steganalysis. Contents: Encompasses data extracted from 1138 cover-stego acoustic pairs (2276 total samples). Features include Reversed-Mel Energy (RME), Derivative-based High-frequency Spectrum (DHS), Wavelet-based Mel-Frequency Cepstral Coefficients (WMC), and Derivative-based MFCC (DMC). This data is pre-formatted for training and evaluating Support Vector Machine (SVM) and FastRealBoostBins (FRBB) classifiers