This study introduces a high-security learning-based optical encryption system using a disordered metasurface (DM) and a double-secure procedure. The DM, with its spin-multiplexing random phase profile, serves as a stable and active speckle generator, providing multiple polarization encryption channels. The plaintext is first encrypted by a security key, which is then superimposed with the plaintext to form a composite signal. This composite signal passes through the DM, generating a speckle pattern that contains both the plaintext and the security key. During decryption, the neural network decodes the speckle pattern using the correct security key and polarization, ensuring two-to-one mapping between the input and output. The system demonstrates excellent stability and decryption efficiency over extended periods in noisy environments, with the DM exhibiting self-recovery capabilities. The proposed method enhances information security and compatibility, making it suitable for practical applications.This study introduces a high-security learning-based optical encryption system using a disordered metasurface (DM) and a double-secure procedure. The DM, with its spin-multiplexing random phase profile, serves as a stable and active speckle generator, providing multiple polarization encryption channels. The plaintext is first encrypted by a security key, which is then superimposed with the plaintext to form a composite signal. This composite signal passes through the DM, generating a speckle pattern that contains both the plaintext and the security key. During decryption, the neural network decodes the speckle pattern using the correct security key and polarization, ensuring two-to-one mapping between the input and output. The system demonstrates excellent stability and decryption efficiency over extended periods in noisy environments, with the DM exhibiting self-recovery capabilities. The proposed method enhances information security and compatibility, making it suitable for practical applications.