Learning-based chosen-plaintext attack on diffractive-imaging-based encryption scheme

The only known approach that can break the diffractive-imaging-based encryption (DIBE) was proposed by Li and Shi in 2015. However, their approach works under the assumption that the phase distribution of the random phase masks (RPMs) is within [0, pi]. In other words, it is no longer effective when such requirement is not fulfilled. In this paper, we propose a universal method, referred to as learning-based chosen-plaintext attack (L-CPA), to break DIBE. The L-CPA enables one to recover the plaintext from the ciphertext by aid of a well-trained artificial neural network (ANN), regardless of the phase distribution of the RPMs. Furthermore, the proposal can be accomplished with no need of knowing the details of the optical arrangement of DIRE. To our best knowledge, this is the first paper that reveals the absolute insecurity of DIBE against CPA. Numerical simulations are presented to demonstrate the effectiveness and feasibility of the proposal.