Malicious Experts Versus the Multiplicative Weights Algorithm in Online Prediction
Malicious Experts Versus the Multiplicative Weights Algorithm in Online Prediction
复制标题
恶意专家与在线预测中的乘法权重算法
DOI：
--
复制
发表时间：
2020
影响因子：
2.5
通讯作者：
Xin Zhang
中科院分区：
文献类型：
--
作者：
Erhan Bayraktar;H. Poor;Xin Zhang
我们考虑由两名专家和一名预测员组成的预测问题。我们假设其中一个专家是诚实的，并且在每一轮中以<inline-formula><tex-math notation="LaTeX">$mu $的</tex-math></inline-formula>概率做出正确的预测。另一个是恶意的，他知道每一轮的真实结果，并做出预测，以最大限度地增加预测者的损失。假设预测者采用经典的乘性权重算法，我们找到恶意专家的值函数的上界<xref rid="deqn5" ref-type="disp-formula">（5）</xref>，以及下界<xref rid="deqn19" ref-type="disp-formula">（19）</xref>。我们的研究结果表明，乘法权重算法不能抵抗恶意专家的腐败。我们还表明，自适应乘法权重算法是渐近最优的预测，因此更能抵抗恶意专家的腐败。
We consider a prediction problem with two experts and a forecaster. We assume that one of the experts is honest and makes correct prediction with probability <inline-formula> <tex-math notation="LaTeX">$mu $ </tex-math></inline-formula> at each round. The other one is malicious, who knows true outcomes at each round and makes predictions in order to maximize the loss of the forecaster. Assuming the forecaster adopts the classical multiplicative weights algorithm, we find an upper bound <xref rid="deqn5" ref-type="disp-formula">(5)</xref> for the value function of the malicious expert, and also a lower bound <xref rid="deqn19" ref-type="disp-formula">(19)</xref>. Our results imply that the multiplicative weights algorithm cannot resist the corruption of malicious experts. We also show that an adaptive multiplicative weights algorithm is asymptotically optimal for the forecaster, and hence more resistant to the corruption of malicious experts.
恶意专家乘法学习系统中的最优对抗策略
DOI：
10.1109/tifs.2021.3052360
发表时间：
2021
影响因子：
6.8
作者：
Etesami, S. Rasoul;Kiyavash, Negar;Leon, Vincent;Poor, H. Vincent
通讯作者：
Poor, H. Vincent