1 /* 2 * Copyright (C) 2023 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 package com.android.cobalt.observations; 18 19 import static com.google.common.base.Preconditions.checkArgument; 20 21 import com.google.cobalt.PrivateIndexObservation; 22 import com.google.cobalt.ReportDefinition; 23 import com.google.common.collect.ImmutableList; 24 import com.google.common.math.BigDecimalMath; 25 26 import java.math.BigDecimal; 27 import java.math.RoundingMode; 28 import java.security.SecureRandom; 29 30 /** Generator for private index observation noise. */ 31 public final class PrivacyGenerator { 32 private final SecureRandom mSecureRandom; 33 PrivacyGenerator(SecureRandom secureRandom)34 public PrivacyGenerator(SecureRandom secureRandom) { 35 this.mSecureRandom = secureRandom; 36 } 37 38 /** 39 * Generates noise for a report. 40 * 41 * @param maxIndex the maximum private index value for the report 42 * @param reportDefinition the privacy-enabled report containing the privacy parameters 43 * @return private indices that include noise according to the report's privacy parameters 44 */ generateNoise( int maxIndex, ReportDefinition reportDefinition)45 ImmutableList<PrivateIndexObservation> generateNoise( 46 int maxIndex, ReportDefinition reportDefinition) { 47 checkArgument(maxIndex >= 0, "maxIndex value cannot be negative"); 48 double lambda = reportDefinition.getShuffledDp().getPoissonMean(); 49 checkArgument(lambda > 0, "poisson_mean must be positive, got %s", lambda); 50 51 // To minimize the number of calls to secureRandom, we draw the total number of 52 // observations, then we distribute those additional observations over the indices. This is 53 // more efficient than drawing from a Poisson distribution once per index when lambda < 1. 54 // We expect lambda << 1 in practice. 55 double lambdaTimesNumIndex = 56 BigDecimalMath.roundToDouble( 57 // (maxIndex + 1) * lambda 58 new BigDecimal(maxIndex) 59 .add(BigDecimal.ONE) 60 .multiply(BigDecimal.valueOf(lambda)), 61 // Set rounding mode to rounding upwards in order to prevent accidentally 62 // using a lower-than-expected noise parameter. 63 RoundingMode.UP); 64 65 int addedOnes = samplePoissonDistribution(lambdaTimesNumIndex); 66 67 ImmutableList.Builder<PrivateIndexObservation> noise = ImmutableList.builder(); 68 for (int i = 0; i < addedOnes; ++i) { 69 noise.add( 70 PrivateIndexObservation.newBuilder() 71 .setIndex(sampleUniformDistribution(maxIndex)) 72 .build()); 73 } 74 75 return noise.build(); 76 } 77 78 /** 79 * Provides samples from the poisson distribution with mean `lambda`. 80 * 81 * <p>Entropy is provided by a SecureRandom generator. Uses the Inverse transform sampling 82 * method, which is good for small lambda, and requires only one secure random generation. We 83 * expect lambda << 1 in practice. For more details, see 84 * https://en.wikipedia.org/wiki/Poisson_distribution#Random_variate_generation. 85 * 86 * @param lambda the poisson mean 87 * @return the sampled number 88 */ samplePoissonDistribution(double lambda)89 private int samplePoissonDistribution(double lambda) { 90 int x = 0; 91 double p = Math.exp(-lambda); 92 double s = p; 93 double u = mSecureRandom.nextDouble(); 94 while (u > s) { 95 x++; 96 p = p * lambda / x; 97 s = s + p; 98 } 99 return x; 100 } 101 102 /** 103 * Provides samples from the uniform distribution over the integers from 0 to `max`, inclusive. 104 * 105 * <p>Entropy is provided by a SecureRandom generator. 106 * 107 * @param max the maximum number to generate. 108 * @return the sampled number 109 */ sampleUniformDistribution(int max)110 private int sampleUniformDistribution(int max) { 111 return mSecureRandom.nextInt(max + 1); 112 } 113 } 114