LSTM-driven chaotic keystream generator for robust medical image encryption

Medical image transmission and storage must be secure due to the rapid growth of digital healthcare systems and telemedicine. Medical images have high data volumes and strong pixel correlations, making conventional encryption algorithms inefficient or computationally expensive. Chaos-based encryption methods are popular due to their sensitivity to initial conditions and strong randomness, but finite precision limits key spaces and short periodicity in chaotic maps. A novel hybrid image encryption scheme is proposed that integrates long short-term memory (LSTM) networks with classical chaotic maps to generate adaptive keystreams using a hash float-seeded chaotic map and statistical features. The chaotic maps provide nonlinearity and high key sensitivity, while the LSTM extends the randomness, eliminates periodicity, and expands the key space to test its robustness. Three variants of the proposed model were experimented with: LSTM + Logistic, LSTM + sine map, and LSTM + Chebyshev map. Results demonstrate that all 3 models achieve strong encryption performance, but LSTM + Logistic map outperforms the other two in terms of randomness quality, noise, and crop attacks. The proposed model achieves 99.62% of NPCR, 39.59% of UACI, and 7.975 of entropy. Compared with conventional chaos methods and GAN encryption schemes, the proposed method provides high security, lightweight computation, and practical deployment.