<p>The increasing reliance on visual data in Internet-of-Things (IoT) and edge intelligence systems necessitates image encryption schemes that offer strong security guarantees while preserving structural flexibility and compatibility with modern learning-based pipelines. This paper presents Quantum Machine Learning (QML)-driven chaotic image encryption framework that integrates a novel Sinusoidal-Polynomial Feedback Chaotic (SPFC) map with a Hybrid Patchwise Quantum Autoencoder (PQAE) and a confusion-diffusion encryption architecture. The proposed PQAE combines a classical teacher autoencoder and a parameterised quantum circuit (PQC) with explicit latent-trash partitioning. The PQAE acts as a dimensionality reduction module that encodes large size colour images into smaller latent representations that can be encrypted efficiently while maintaining high-fidelity reconstruction. Moreover, the designed SPFC map is employed as a highly secure key generator and is rigorously validated using bifurcation analysis, Lyapunov exponent and Kolmogorov entropy measurements, and the NIST SP 800-22 statistical test suite, achieving a maximum Lyapunov exponent of approximately 9.7, Kolmogorov entropy of approximately 4.44, and a 99.9% NIST pass rate. The encoded latent images are encrypted using a two-stage permutation process, dynamic multi-S-box substitution, chaotic bit shifts, and XOR-based diffusion driven by SPFC-derived keys. Comprehensive security evaluation demonstrates near-ideal security statistics, including entropy values up to 7.9971, negligible pixel correlation, and strong resistance to differential attacks with NPCR exceeding <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$99.65\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>99.65</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> and UACI in the range of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$33.46\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>33.46</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>–<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$33.48\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>33.48</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> across standard benchmark images. Robustness experiments further confirm reliable decryption under occlusion (20%–50%) attacks and salt-and-pepper noise perturbations of up to 20%.</p>

Quantum-assisted image encryption via patchwise quantum autoencoder and novel sinusoidal-polynomial feedback chaotic map

期刊论文
发表日期：

2026年4月17日

查看全文

Jawad Ahmad

摘要

The increasing reliance on visual data in Internet-of-Things (IoT) and edge intelligence systems necessitates image encryption schemes that offer strong security guarantees while preserving structural flexibility and compatibility with modern learning-based pipelines. This paper presents Quantum Machine Learning (QML)-driven chaotic image encryption framework that integrates a novel Sinusoidal-Polynomial Feedback Chaotic (SPFC) map with a Hybrid Patchwise Quantum Autoencoder (PQAE) and a confusion-diffusion encryption architecture. The proposed PQAE combines a classical teacher autoencoder and a parameterised quantum circuit (PQC) with explicit latent-trash partitioning. The PQAE acts as a dimensionality reduction module that encodes large size colour images into smaller latent representations that can be encrypted efficiently while maintaining high-fidelity reconstruction. Moreover, the designed SPFC map is employed as a highly secure key generator and is rigorously validated using bifurcation analysis, Lyapunov exponent and Kolmogorov entropy measurements, and the NIST SP 800-22 statistical test suite, achieving a maximum Lyapunov exponent of approximately 9.7, Kolmogorov entropy of approximately 4.44, and a 99.9% NIST pass rate. The encoded latent images are encrypted using a two-stage permutation process, dynamic multi-S-box substitution, chaotic bit shifts, and XOR-based diffusion driven by SPFC-derived keys. Comprehensive security evaluation demonstrates near-ideal security statistics, including entropy values up to 7.9971, negligible pixel correlation, and strong resistance to differential attacks with NPCR exceeding $99.65\%$ $99.65 %$ and UACI in the range of $33.46\%$ $33.46 %$ – $33.48\%$ $33.48 %$ across standard benchmark images. Robustness experiments further confirm reliable decryption under occlusion (20%–50%) attacks and salt-and-pepper noise perturbations of up to 20%.