Stochastic Yet Precise: Memristor Crossbar Arrays Enabling Robust In-Memory Computing, Hardware Security, and Generative Adversarial Network

Here, we propose a multifunctional platform based on an oxide-based RRAM crossbar array that integrates in-memory computing, hardware security, and generative AI functionalities. By employing a pulse programming algorithm, we achieve 5-bit current state control, enabling stable and accurate analog conductance programming for vector–matrix multiplication (VMM) without external data movement. The intrinsic variability is further harnessed as a high-entropy source to construct a physical unclonable function (PUF) with fast bit generation throughput, ensuring robust hardware-based security. Beyond computation and security, the random bitstreams generated by the PUF are further employed as seeds for generative AI models, enabling the synthesis of structurally diverse and perceptually realistic iris images that show superior quantitative performance in multi-scale structural similarity (MS-SSIM) and learned perceptual image patch similarity (LPIPS) compared to those produced using conventional software noise. These findings highlight how RRAM crossbar arrays, despite their inherent stochasticity, can be precisely controlled to enable reliable VMM, secure key generation, and high-fidelity data synthesis. This multifunctional approach underscores the transformative potential of RRAM architectures as foundational building blocks for next-generation intelligent and secure edge systems.