DGA-based botnet detection toward imbalanced multiclass learning

Botnets based on the Domain Generation Algorithm (DGA) mechanism pose great challenges to main current detection methods because of their strong concealment and robustness. However, complexity DGA family imbalance samples continue impede research detection. In existing work, sample size each is regarded as most important determinant resampling proportion; thus, differences in characteristics various are ignored, optimal effect not achieved. this paper, a Long Short-Term Memory-based Property Quantity Dependent Optimization (LSTM.PQDO) method proposed. This takes advantage LSTM automatically mine comprehensive features domain names. It iterates proportion with solution consideration original number heuristically search for better around initial right direction; dynamic optimization realized. The experimental results show that LSTM.PQDO can achieve performance compared models overcome difficulties unbalanced datasets; moreover, it function reference tasks similar scenarios.