Abstract: The combination of Distributed Face Recognition System (DFRS) based on deep learning and secure computation has compatibility and performance problems. The compatibility problem is manifested in the difference between the nonlinear function in deep learning and the types of operations supported by secure computation. The performance problem is that the combination of inefficient secure computation and high-overhead deep learning will lead to excessive response time of DFRS. Aiming at the compatibility problem and taking into account the nonlinearity, a privacy-protected DFRS based on the combination of garbled circuit and deep learning is proposed to solve the problem of computing compatibility; For the performance problem, a reform scheme of the architecture is proposed. The network layer is split into two parts, and the more computationally intensive part is deployed on the client in the form of efficient plaintext calculation to improve the efficiency of DFRS. Theoretical analysis shows that the new scheme has both the functional advantages of non-linear calculation and security protection. The prototype experiment shows that the new architecture reduces the message complexity of the system from nearly 1998.206 MB to nearly 60.591 MB, the pure computing time from 17.742 s to 0.644 s, and the system response time from 177.125 s to 5.751 s under 100 MB bandwidth.