The energy cost of network security: A hardware vs. software comparison

The increasing network speeds, number of attacks, and need for energy efficiency are pushing software-based network security to the limit. A common kind of threat is probing attacks, in which an attacker tries to find vulnerabilities by sending many probe packets to a target machine. In this paper, we evaluate three machine learning classifiers (Decision Tree, Naive Bayes, and k-Nearest Neighbors), implemented in hardware and software, for the detection of probing attacks. We present detailed results showing the tradeoffs between energy consumption, throughput, and accuracy of the three classifiers. The fastest hardware implementation is 926 times as fast as its software counterpart, and its energy consumption per classification is 0.05% that of the software version.