2016/12/2(五) 14:20 - 陳伯寧 教授 (交大電機系) - Optimal Byzantine Attack for Distributed Inference
Topic： Optimal Byzantine Attack for Distributed Inference
Date&Time：2016/12/2 (五) 14:20
Speaker：陳伯寧 教授 (交大電機系)
Location：清華大學台達館 R216
Abstract：
Robustness of distributed inference against Byzantine attacks is one of the important considerations in a wireless sensor network (WSN) design. In this subject, instead of working on system robustness from the standpoint of a system designer, we investigate it from that of a Byzantine attacker. Under the system setting that each local sensor sends M-ary data to the fusion center, an optimal Byzantine attack policy that can blind the system with the minimum fraction of compromised sensors is derived under the assumption that the Byzantine adversary has complete knowledge of the statistics of local quantization outputs. Closed-form expressions for the minimum fraction of compromised sensors 內置圖片 and the 內置圖片-achieving Byzantine transition probability are also obtained. Our results indicate that the statistics of the quantization outputs of local sensors is essential for an attacker to optimize its attack and, therefore, it should be protected well.
Date&Time：2016/12/2 (五) 14:20
Speaker：陳伯寧 教授 (交大電機系)
Location：清華大學台達館 R216
Abstract：
Robustness of distributed inference against Byzantine attacks is one of the important considerations in a wireless sensor network (WSN) design. In this subject, instead of working on system robustness from the standpoint of a system designer, we investigate it from that of a Byzantine attacker. Under the system setting that each local sensor sends M-ary data to the fusion center, an optimal Byzantine attack policy that can blind the system with the minimum fraction of compromised sensors is derived under the assumption that the Byzantine adversary has complete knowledge of the statistics of local quantization outputs. Closed-form expressions for the minimum fraction of compromised sensors 內置圖片 and the 內置圖片-achieving Byzantine transition probability are also obtained. Our results indicate that the statistics of the quantization outputs of local sensors is essential for an attacker to optimize its attack and, therefore, it should be protected well.