Hybrid Quantum Deep Learning with Differential Privacy for Botnet DGA Detection

In the DNS query-based botnet domain generation algorithm (DGA) detection, one might argue that domain names in DNS query data might disclose sensitive information related to browsing histories. User privacy preservation is important in the current personal data protection (PDP) era. This paper proposed implementing the differential privacy approach to the hybrid quantum deep learning model for botnet DGA detection. The proposed model consists of traditional deep learning layers and a quantum layer by combining angle embedding and random layer circuits from the Pennylane framework. We used ten botnet DGA datasets: Conficker, Cryptolocker, Goz, Matsnu, New_Goz, Pushdo, Ramdo, and Rovnix. We conducted experiments with considering noise models of eight IBM quantum devices: (ibmq_5_yorktown, ibmq_armonk, ibmq_athens, ibmq_belem, ibmq_lima, ibmq_quito, ibmq_santiago, and ibmqx2). We found that our proposed hybrid quantum model delivers a satisfactory performance (92.4% of maximum accuracy), superior to the classical deep learning counterpart. However, the hyperparameters of the differential privacy implementations (l2_norm_clip, noise_multiplier, microbatches, and learning_rate) still need to be tuned to improve the privacy guarantee of our proposed models.