DISTRIBUTED PROCESSING OF DEEP LEARNING INFERENCE MODELS FOR MALICIOUS URL DETECTION

초록

Stacking used for improving the accuracy of deep learning models incurs a long inference time due to its high complexity, and it further increases the latency significantly in data stream environments. In particular, the long latency is a severe problem in detecting malicious URLs since the real-time detection is a critical requirement. In this paper, we propose a distributed processing technique that efficiently processes a stacking-based inference model of detecting malicious URLs in data stream environments. Distributed algorithms may vary greatly in processing performance depending on their configurations, so we propose four different configurations: Independent Stacking, Sequential Stacking, Semi-Sequential Stacking, and Stepwise-Independent Stacking. We then evaluate the four configurations by comparing the latency and resource usage occurring in processing URL streams. Experimental results show that Stepwise-Independent Stacking, which has the property of both independent and sequential executions, is the most efficient configuration by providing the shortest latency. ICIC International © 2022.

키워드