Implementing lightweight context-aware N-variant systems with deep learning for practical system defense

Aiming at the problem that software and hardware differentiation easily leads to inconsistent decision-making results in N-variant systems, causing false positive phenomena to be mistaken for cyberattacks, a lightweight context-aware N-variant decision-making method based on deep learning is proposed. By constructing a minimalist autoencoder-decoder deep learning model and using unsupervised learning, deep semantic features of diversified normal response data output by different variants are mined, their statistical regularity is analyzed and summarized, and the detection domain of normal response data is accurately acquired. The false positive phenomenon is solved by designing a training mechanism based on offline learning-online decision-making linkage adapted to real application scenarios and a feedback optimization mechanism, thereby accurately detecting cyberattacks. To further verify the effectiveness of our method, it was deployed and run in two real applications, the government cadre performance evaluation management platform and the power system. Experimental results show that the detection performance of the proposed method is significantly superior to mainstream N-variant decision-making methods, and the average prediction accuracy is improved by 14.89%.