Abstract
This study investigates the robustness of a chaos-based audio encryption scheme against learning-assisted cryptanalysis using Echo State Networks (ESNs). The encryption pipeline is driven by a chaotic key stream generated from the Rössler attractor. The plaintext audio is segmented into fixed-length blocks, permuted in time, scrambled in the frequency domain via a conjugate-symmetric permutation, and further obscured by a structured frequency-domain phase mask, yielding a noise-like ciphertext. A partial-leakage threat model is considered in which an eavesdropper (Eve) knows the algorithmic structure and block order, observes only a fraction of the frequency-permutation mappings, and receives noisy phase samples on a randomly selected subset of blocks (random missing leakage). Here, "random missing leakage" refers to a setting in which phase observations are available only for a randomly selected subset of blocks, while the remaining blocks contain no phase information. To exploit the structured phase model, Eve trains an ESN only on leakage-observed blocks and imputes latent phase-basis coefficients for non-observed blocks, enabling approximate decryption with the estimated permutation and reconstructed phase mask. Reconstruction quality is evaluated using intelligibility- and error-oriented metrics (STOI, SNR, CORR, and NMSE), reported both overall and separately on leakage-observed and leakage-missing segments. The results reveal a non-linear transition in adversarial recoverability as permutation leakage increases, and highlight the coupled roles of frequency-bin alignment and phase-model learnability in determining the attack effectiveness.