Vulnerability to machine learning attacks of optical encryption based on diffractive imaging

• Published in: Optics and lasers in engineering 2020-02, Vol.125, p.105858, Article 105858
• Main Authors: Zhou, Lina, Xiao, Yin, Chen, Wen
• Format: Article
• Language: English
• Subjects: Diffractive imaging; Experimental demonstration; Machine learning; Vulnerability detection
• ISSN: 0143-8166; 1873-0302
• DOI: 10.1016/j.optlaseng.2019.105858

•It is experimentally demonstrated that optical encryption based on diffractive imaging is vulnerable to machine learning attacks.•Optical encryption based on diffractive imaging cannot withstand the proposed machine learning attacks.•The trained learning model can extract unknown plaintexts from the given ciphertexts in real time without the direct retrieval or estimate of various different optical encryption keys.•The universality of the trained learning model is also illustrated, and it is verified that the machine learning model trained by using a database is robust to be used for attacking different databases.•The proposed machine learning attacks would urge the further investigation of optical encryption schemes to enhance their security.•The proposed machine learning attacks could hold the promise for further developments of the cryptanalysis of optical encryption. In this paper, we experimentally demonstrate for the first time to our knowledge that optical encryption based on diffractive imaging is vulnerable to the attacks using learning methods. Using machine learning attack, an opponent is capable to retrieve unknown plaintexts from the given ciphertexts. The proposed method adopts end-to-end learning to extract a superior mapping relationship between the ciphertexts and the plaintexts. Without a direct retrieval or estimate of optical encryption keys, an unauthorised user can extract unknown plaintexts from the given ciphertexts by using the trained learning models. Simulations and optical experimental results demonstrate that the proposed learning method is feasible and effective to analyze the vulnerability of optical encryption schemes. The universality of the trained learning model is also illustrated, and it is verified that the machine learning model trained by using a database is robust to be used for attacking different databases. Compared with conventional cryptanalytic methods, the proposed machine learning attacks can retrieve unknown plaintexts from the given ciphertexts using the trained learning models without a direct usage of various different optical encryption keys, which provides a different strategy for the cryptanalysis of optical encryption systems. @ Elsevier, 2019.