arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2107.08461cs.LGcs.CRstat.ML
收藏

精确性、保密性和可靠性的差分私有贝叶斯神经网络

Differentially Private Bayesian Neural Networks on Accuracy, Privacy and Reliability

University of Pennsylvania(宾夕法尼亚大学)

Qiyiwen Zhang, Zhiqi Bu, Kan Chen, Qi Long

摘要:

贝叶斯神经网络(BNN)允许预测中的不确定性量化,提供了比常规神经网络更大的优势,这在差分隐私(DP)框架中还没有被探索过。我们通过利用贝叶斯深度学习和隐私会计的最新发展来填补这一重要空白,从而对BNN中隐私和准确性之间的权衡进行更精确的分析。我们提出了三种DP-BNN,它们以不同的方式表征同一网络架构的权重不确定性,即DP-SGLD(通过噪声梯度方法),DP-BBP(通过改变感兴趣的参数)和DP-MC Dropout(通过模型架构)。有趣的是,我们展示了DP-SGD和DP-SGLD之间的新等价性,这意味着一些非贝叶斯DP训练自然允许不确定性量化。然而,超参数,如学习率和批量大小,可以在DP-SGD和DP-SGLD中具有不同甚至相反的效果。进行了大量的实验,比较DP-BNN,在隐私保证,预测精度,不确定性量化,校准,计算速度和网络架构的推广。因此,我们观察到一个新的权衡之间的隐私和可靠性。与非DP和非贝叶斯方法相比,DP-SGLD在强隐私保证下具有非常高的准确性,证明了DP-BNN在现实任务中的巨大潜力。

英文摘要:

Bayesian neural network (BNN) allows for uncertainty quantification in prediction, offering an advantage over regular neural networks that has not been explored in the differential privacy (DP) framework. We fill this important gap by leveraging recent development in Bayesian deep learning and privacy accounting to offer a more precise analysis of the trade-off between privacy and accuracy in BNN. We propose three DP-BNNs that characterize the weight uncertainty for the same network architecture in distinct ways, namely DP-SGLD (via the noisy gradient method), DP-BBP (via changing the parameters of interest) and DP-MC Dropout (via the model architecture). Interestingly, we show a new equivalence between DP-SGD and DP-SGLD, implying that some non-Bayesian DP training naturally allows for uncertainty quantification. However, the hyperparameters such as learning rate and batch size, can have different or even opposite effects in DP-SGD and DP-SGLD. Extensive experiments are conducted to compare DP-BNNs, in terms of privacy guarantee, prediction accuracy, uncertainty quantification, calibration, computation speed, and generalizability to network architecture. As a result, we observe a new tradeoff between the privacy and the reliability. When compared to non-DP and non-Bayesian approaches, DP-SGLD is remarkably accurate under strong privacy guarantee, demonstrating the great potential of DP-BNN in real-world tasks.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑