差分隐私与联邦学习：分布式机器学习的隐私保护之道

\u5f15\u8a00

\u5728\u6570\u636e\u6210\u4e3a\u65b0\u65f6\u4ee3\u77f3\u6cb9\u7684\u4eca\u5929\uff0c\u5982\u4f55\u5728\"\u5229\u7528\u6570\u636e\"\u4e0e\"\u4fdd\u62a4\u9690\u79c1\"\u4e4b\u95f4\u627e\u5230\u5e73\u8861\uff0c\u662f\u6bcf\u4e00\u4f4d\u6570\u636e\u79d1\u5b66\u5bb6\u548c\u5de5\u7a0b\u5e08\u5fc5\u987b\u9762\u5bf9\u7684\u8bfe\u9898\u3002\u5dee\u5206\u9690\u79c1\uff08Differential Privacy, DP\uff09\u63d0\u4f9b\u4e86\u4e00\u79cd\u53ef\u8bc1\u660e\u7684\u3001\u6570\u5b66\u4e0a\u4e25\u683c\u7684\u9690\u79c1\u4fdd\u62a4\u6846\u67b6\uff0c\u800c\u8054\u90a6\u5b66\u4e60\uff08Federated Learning, FL\uff09\u5219\u8ba9\u6570\u636e\"\u53ef\u7528\u4e0d\u53ef\u89c1\"\u4ece\u613f\u666f\u8d70\u5411\u5de5\u7a0b\u5b9e\u8df5\u3002\u8fd9\u4e24\u9879\u6280\u672f\u7684\u7ed3\u5408\u2014\u2014\u5dee\u5206\u9690\u79c1\u8054\u90a6\u5b66\u4e60\uff08DP-FL\uff09\uff0c\u6b63\u5728\u6210\u4e3a\u91d1\u878d\u3001\u533b\u7597\u3001\u653f\u52a1\u7b49\u654f\u611f\u6570\u636e\u9886\u57df\u7684\u57fa\u7840\u8bbe\u65bd\u3002

\u672c\u6587\u5c06\u4ece\u5dee\u5206\u9690\u79c1\u7684\u6570\u5b66\u5b9a\u4e49\u51fa\u53d1\uff0c\u9010\u6b65\u63a8\u5bfc Laplace \u673a\u5236\u4e0e Gaussian \u673a\u5236\u7684\u8bbe\u8ba1\u539f\u7406\uff0c\u6df1\u5165\u5206\u6790\u8054\u90a6\u5b66\u4e60\u7684\u901a\u4fe1\u4f18\u5316\u4e0e\u805a\u5408\u7b56\u7565\uff0c\u6700\u7ec8\u5c55\u793a\u4e00\u4e2a\u5b8c\u6574\u7684 DP-FL \u7cfb\u7edf\u5b9e\u73b0\u3002

\u4e00\u3001\u5dee\u5206\u9690\u79c1\u7684\u6570\u5b66\u57fa\u7840

1.1 \u5f62\u5f0f\u5316\u5b9a\u4e49

\u5dee\u5206\u9690\u79c1\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u65e0\u8bba\u67d0\u4e2a\u4e2a\u4f53\u7684\u6570\u636e\u662f\u5426\u5b58\u5728\u4e8e\u6570\u636e\u5e93\u4e2d\uff0c\u67e5\u8be2\u7ed3\u679c\u7684\u6982\u7387\u5206\u5e03\u51e0\u4e4e\u76f8\u540c\u3002\u5f62\u5f0f\u5316\u5730\uff0c\u5bf9\u4e8e\u4efb\u610f\u4e24\u4e2a\u76f8\u90bb\u6570\u636e\u96c6 D\u2081 \u548c D\u2082\uff08\u4ec5\u5dee\u4e00\u6761\u8bb0\u5f55\uff09\uff0c\u4e00\u4e2a\u968f\u673a\u7b97\u6cd5 M \u6ee1\u8db3 (\u03b5, \u03b4)-\u5dee\u5206\u9690\u79c1\uff0c\u5f53\u4e14\u4ec5\u5f53\u5bf9\u6240\u6709\u53ef\u80fd\u7684\u8f93\u51fa\u5b50\u96c6 S\uff1a

Pr[M(D\u2081) \u2208 S] \u2264 e\u1d49 \u00b7 Pr[M(D\u2082) \u2208 S]   \u03b4

\u5f53 \u03b4 = 0 \u65f6\uff0c\u79f0\u4e3a\u7eaf\u5dee\u5206\u9690\u79c1\uff08\u03b5-DP\uff09\uff1b\u03b4