计算机科学 ›› 2023, Vol. 50 ›› Issue (11): 122-131.doi: 10.11896/jsjkx.220900169
倪洪杰1, 刘嘉威1, 郑海斌1,2, 陈奕芃1, 陈晋音1,2
NI Hongjie1, LIU Jiawei1, ZHENG Haibin1,2, CHEN Yipeng1, CHEN Jinyin1,2
摘要: 随着深度学习的广泛应用,研究者在关注模型分类性能的同时,还需要关注模型的决策是否公平可信。存在决策偏见的深度模型会造成极大的负面影响,因此如何维持深度模型的分类正确率,同时提高模型的决策公平至关重要。目前已有工作提出了较多方法,用于改善模型的个体公平,但是这些方法仍然在去偏效果、去偏后模型可用性、去偏效率等方面存在缺陷。为此,文中分析了深度模型存在个体偏见时神经元异常激活现象,提出了一种基于偏见神经元抑制的模型去偏方法NeuronSup,具有显著降低个体偏见、对主任务性能影响小、时间复杂度低等优势。具体而言,首先根据深度模型部分神经元由于个体偏见而产生异常激活的现象提出了偏见神经元的概念。然后,利用歧视样本对查找深度模型中的偏见神经元,通过抑制偏见神经元的异常激活大幅降低深度模型的个体偏见,并且根据每个神经元的最大权重边确定主任务性能神经元,通过保持深度模型的主任务性能神经元参数不变,来减小去偏操作对深度模型分类性能造成的影响。因为 NeuronSup只对深度模型中的特定神经元进行去偏操作,所以时间复杂度更低,效率更高。最后,在3个真实数据集的6种敏感属性上开展去偏实验,与5种对比算法相比,NeuronSup将个体公平指标THEMIS降低了50%以上,同时使去偏操作对深度模型分类准确率的影响降低到3%以内,验证了NeuronSup在保证深度模型分类能力的情况下降低个体偏见的有效性。
中图分类号:
[1]HARALICK R M,SHANMUGAM K,DINSTEIN I.TexturalFeatures for Image Classification [J].IEEE Transactions on Systems,Man,and Cybernetics,1973,3(6):610-621. [2]CHAR D S,SHAH N H,MAGNUS D.Implementing MachineLearning in Health Care-Addressing Ethical Challenges[J].New England Journal of Medicine,2018,378(11):981-983. [3]BRENNAN T,DIETERICH W,EHRET B.Evaluating thePredictive Validity of the Compas Risk and Needs Assessment System[J].Criminal Justice & Behavior,2008,36(1):21-40. [4]LIU L T,DEAN S,ROLF E,et al.Delayed impact of fair machine learning[C]//International Conference on Machine Lear-ning.PMLR,2018:3150-3158. [5]WADSWORTH C,VERA F,PIECH C.Achieving fairnessthrough adversarial learning:an application to recidivism prediction [J].arXiv:1807.00199,2018. [6]LICHMAN M.UCI machine learning repository[OL].http://archive.ics.uci.edu/ml. [7]HARDT M,PRICE E,SREBRO N.Equality of opportunity in supervised learning [J].arXiv:1610.02413,2016. [8]FELDMAN M,FRIEDLER S A,MOELLER J,et al.Certifying and removing disparate impact[C]//Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Disco-very and Data Mining.2015:259-268. [9]CHAKRABORTY J,MAJUMDER S,MENZIES T.Bias in machine learning software:why? how? what to do?[C]//Procee-dings of the 29th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering.2021:429-440. [10]PENG K,CHAKRABORTY J,MENZIES T.xFAIR:BetterFairness via Model-based Rebalancing of Protected Attributes[J].arXiv:2110.01109,2021. [11]CALMON F P,WEI D,VINZAMURI B,et al.Optimized pre-processing for discrimination prevention[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems.2017:3995-4004. [12]KAMIRAN F,CALDERS T.Data preprocessing techniques for classification without discrimination[J].Knowledge and Information Systems,2012,33(1):1-33. [13]FELDMAN M,FRIEDLER S A,MOELLER J,et al.Certifying and removing disparate impact[C]//Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Disco-very and Data Mining.2015:259-268. [14]YUROCHKIN M,SUN Y.Sensei:Sensitive set invariance forenforcing individual fairness[J].arXiv:2006.14168,2020. [15]YUROCHKIN M,BOWER A,SUN Y.Training individuallyfair ML models with sensitive subspace robustness[J].arXiv:1907.00020,2019. [16]LOHIA P K,RAMAMURTHY K N,BHIDE M,et al.Bias mi-tigation post-processing for individual and group fairness[C]//2019 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP 2019).IEEE,2019:2847-2851. [17]KIM M P,GHORBANI A,ZOU J.Multiaccuracy:Black-boxpost-processing for fairness in classification[C]//Proceedings of the 2019 AAAI/ACM Conference on AI,Ethics,and Society.2019:247-254. [18]ZEMEL R,WU Y,SWERSKY K,et al.Learning fair representations[C]//International Conference on Machine Learning.PMLR,2013:325-333. [19]SATTIGERI P,HOFFMAN S C,CHENTHAMARAKSHANV,et al.Fairness GAN:Generating datasets with fairness pro-perties using a generative adversarial network[J].IBM Journal of Research and Development,2019,63(4/5):3:1-3:9. [20]ZHANG B H,LEMOINE B,MITCHELL M.Mitigating un-wanted biases with adversarial learning[C]//Proceedings of the 2018 AAAI/ACM Conference on AI,Ethics,and Society.2018:335-340. [21]XU D,YUAN S,ZHANG L,et al.Fairgan:Fairness-aware gene-rative adversarial networks[C]//2018 IEEE International Conference on Big Data(Big Data).IEEE,2018:570-575. [22]BEUTEL A,CHEN J,ZHAO Z,et al.Data decisions and theoretical implications when adversarially learning fair representations[J].arXiv:1707.00075,2017. [23]KAMISHIMA T,AKAHO S,SAKUMA J.Fairness-awarelearning through regularization approach[C]//2011 IEEE 11th International Conference on Data Mining Workshops.IEEE,2011:643-650. [24]ZHANG P,WANG J,SUN J,et al.White-box fairness testing through adversarial sampling[C]//Proceedings of the ACM/IEEE 42nd International Conference on Software Engineering.2020:949-960. [25]UDESHI S,ARORA P,CHATTOPADHYAY S.Automated directed fairness testing[C]//Proceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering.2018:98-108. [26]GALHOTRA S,BRUN Y,MELIOU A.Fairness testing:tes-ting software for discrimination[C]//Proceedings of the 2017 11th Joint Meeting on Foundations of Software Engineering.2017:498-510. [27]KAMIRAN F,MANSHA S,KARIM A,et al.Exploiting reject option in classification for social discrimination control[J].Information Sciences,2018,425:18-33. [28]LIU Z,LI J,SHEN Z,et al.Learning efficient convolutional net-works through network slimming[C]//Proceedings of the IEEE International Conference on Computer Vision.2017:2736-2744. |
|