WIP: Differential privacy for aggregates - add dp_laplace and dp_gaussian built-in functions with budget accounting

pom.xml

@@ -408,6 +408,7 @@
 				<artifactId>maven-surefire-plugin</artifactId>
 				<version>${maven-surefire-plugin.version}</version>
 				<configuration>
+					<reportFormat>plain</reportFormat>
 					<skipTests>${maven.test.skip}</skipTests>
 					<parallel>${test-parallel}</parallel>
 					<threadCount>${test-threadCount}</threadCount>

src/main/java/org/apache/sysds/common/Builtins.java

@@ -116,6 +116,8 @@ public enum Builtins {
 	DECISIONTREEPREDICT("decisionTreePredict", true),
 	DECOMPRESS("decompress", false),
 	DEDUP("dedup", true),
+	DP_LAPLACE("dp_laplace", false),
+	DP_GAUSSIAN("dp_gaussian", false),
 	DEEPWALK("deepWalk", true),
 	DET("det", false),
 	DETECTSCHEMA("detectSchema", false),

src/main/java/org/apache/sysds/common/InstructionType.java

@@ -63,6 +63,7 @@ public enum InstructionType {
 	MMChain,
 	Union,
 	EINSUM,
+	DPBuiltin,
 
 	//SP Types
 	MAPMM,

src/main/java/org/apache/sysds/common/Opcodes.java

@@ -194,6 +194,10 @@ public enum Opcodes {
 	LIST("list", InstructionType.BuiltinNary),
 	EINSUM("einsum", InstructionType.BuiltinNary),
 
+	//DP built-in functions
+	DP_LAPLACE("dp_laplace", InstructionType.DPBuiltin),
+	DP_GAUSSIAN("dp_gaussian", InstructionType.DPBuiltin),
+
 	//Parametrized builtin functions
 	AUTODIFF("autoDiff", InstructionType.ParameterizedBuiltin),
 	CONTAINS("contains", InstructionType.ParameterizedBuiltin),

src/main/java/org/apache/sysds/common/Types.java

@@ -806,7 +806,8 @@ public static ReOrgOp valueOfByOpcode(String opcode) {
 
 	/** Parameterized operations that require named variable arguments */
 	public enum ParamBuiltinOp {
-		AUTODIFF, CDF, CONTAINS, INVALID, INVCDF, GROUPEDAGG, RMEMPTY, REPLACE, REXPAND,
+		AUTODIFF, CDF, CONTAINS, DP_LAPLACE, DP_GAUSSIAN, INVALID, INVCDF,
+		GROUPEDAGG, RMEMPTY, REPLACE, REXPAND,
 		LOWER_TRI, UPPER_TRI,
 		TRANSFORMAPPLY, TRANSFORMDECODE, TRANSFORMCOLMAP, TRANSFORMMETA,
 		TOKENIZE, TOSTRING, LIST, PARAMSERV

src/main/java/org/apache/sysds/hops/ParameterizedBuiltinOp.java

@@ -182,7 +182,7 @@ public Lop constructLops()
 			}
 			case CONTAINS:
 			case CDF:
-			case INVCDF: 
+			case INVCDF:
 			case REPLACE:
 			case LOWER_TRI:
 			case UPPER_TRI:
@@ -194,7 +194,9 @@ public Lop constructLops()
 			case TOSTRING:
 			case PARAMSERV:
 			case LIST:
-			case AUTODIFF:{
+			case AUTODIFF:
+			case DP_LAPLACE:
+			case DP_GAUSSIAN:{
 				ParameterizedBuiltin pbilop = new ParameterizedBuiltin(
 					inputlops, _op, getDataType(), getValueType(), et);
 				if( isMultiThreadedOpType() )
@@ -688,7 +690,11 @@ else if( _op == ParamBuiltinOp.TRANSFORMAPPLY ) {
 				return new MatrixCharacteristics(dc.getRows(), dc.getCols(), -1, dc.getLength());
 			}
 		}
-
+		else if( _op == ParamBuiltinOp.DP_LAPLACE || _op == ParamBuiltinOp.DP_GAUSSIAN ) {
+			if( dc.dimsKnown() )
+				ret = new MatrixCharacteristics(dc.getRows(), dc.getCols(), -1, dc.getLength());
+		}
+
 		return ret;
 	}
 	@Override
@@ -758,7 +764,8 @@ && getTargetHop().areDimsBelowThreshold() ) {
 		if (_op == ParamBuiltinOp.TRANSFORMCOLMAP || _op == ParamBuiltinOp.TRANSFORMMETA
 				|| _op == ParamBuiltinOp.TOSTRING || _op == ParamBuiltinOp.LIST
 				|| _op == ParamBuiltinOp.CDF || _op == ParamBuiltinOp.INVCDF
-				|| _op == ParamBuiltinOp.PARAMSERV) {
+				|| _op == ParamBuiltinOp.PARAMSERV
+				|| _op == ParamBuiltinOp.DP_LAPLACE || _op == ParamBuiltinOp.DP_GAUSSIAN) {
 			_etype = ExecType.CP;
 		}
 

src/main/java/org/apache/sysds/lops/ParameterizedBuiltin.java

@@ -204,7 +204,19 @@ public String getInstructions(String output)
 				compileGenericParamMap(sb, _inputParams);
 				break;
 			}
-
+			case DP_LAPLACE: {
+				sb.append(Opcodes.DP_LAPLACE);
+				sb.append(OPERAND_DELIMITOR);
+				compileGenericParamMap(sb, _inputParams);
+				break;
+			}
+			case DP_GAUSSIAN: {
+				sb.append(Opcodes.DP_GAUSSIAN);
+				sb.append(OPERAND_DELIMITOR);
+				compileGenericParamMap(sb, _inputParams);
+				break;
+			}
+
 			default:
 				throw new LopsException(this.printErrorLocation() + "In ParameterizedBuiltin Lop, Unknown operation: " + _operation);
 		}

src/main/java/org/apache/sysds/parser/BuiltinFunctionExpression.java

@@ -2005,6 +2005,31 @@ else if(this.getOpCode() == Builtins.MAX_POOL || this.getOpCode() == Builtins.AV
 			}
 			else 
 				raiseValidateError("Local instruction not allowed in dml script");
+		case DP_LAPLACE: {
+			checkNumParameters(3);
+			checkMatrixParam(getFirstExpr());
+			checkScalarParam(getSecondExpr());
+			checkScalarParam(getThirdExpr());
+			output.setDataType(DataType.MATRIX);
+			output.setValueType(ValueType.FP64);
+			output.setDimensions(
+				getFirstExpr().getOutput().getDim1(),
+				getFirstExpr().getOutput().getDim2());
+			break;
+		}
+		case DP_GAUSSIAN: {
+			checkNumParameters(4);
+			checkMatrixParam(getFirstExpr());
+			checkScalarParam(getSecondExpr());
+			checkScalarParam(getThirdExpr());
+			checkScalarParam(getFourthExpr());
+			output.setDataType(DataType.MATRIX);
+			output.setValueType(ValueType.FP64);
+			output.setDimensions(
+				getFirstExpr().getOutput().getDim1(),
+				getFirstExpr().getOutput().getDim2());
+			break;
+		}
 		case COMPRESS:
 		case DECOMPRESS:
 			if(OptimizerUtils.ALLOW_SCRIPT_LEVEL_COMPRESS_COMMAND){

src/main/java/org/apache/sysds/parser/DMLTranslator.java

@@ -2310,6 +2310,10 @@ private Hop processBuiltinFunctionExpression(BuiltinFunctionExpression source, D
 		if (source.getThirdExpr() != null) {
 			expr3 = processExpression(source.getThirdExpr(), null, hops);
 		}
+		Hop expr4 = null;
+		if (source.getFourthExpr() != null) {
+			expr4 = processExpression(source.getFourthExpr(), null, hops);
+		}
 
 		Hop currBuiltinOp = null;
 		target = (target == null) ? createTarget(source) : target;
@@ -2589,6 +2593,25 @@ else if ( sop.equalsIgnoreCase(Opcodes.NOTEQUAL.toString()) )
 		case DECOMPRESS:
 			currBuiltinOp = new UnaryOp(target.getName(), target.getDataType(), ValueType.FP64, OpOp1.DECOMPRESS, expr);
 			break;
+		case DP_LAPLACE: {
+			LinkedHashMap<String, Hop> dpLaplaceParams = new LinkedHashMap<>();
+			dpLaplaceParams.put("target", expr);
+			dpLaplaceParams.put("sensitivity", expr2);
+			dpLaplaceParams.put("epsilon", expr3);
+			currBuiltinOp = new ParameterizedBuiltinOp(target.getName(), DataType.MATRIX,
+				ValueType.FP64, ParamBuiltinOp.DP_LAPLACE, dpLaplaceParams);
+			break;
+		}
+		case DP_GAUSSIAN: {
+			LinkedHashMap<String, Hop> dpGaussianParams = new LinkedHashMap<>();
+			dpGaussianParams.put("target", expr);
+			dpGaussianParams.put("sensitivity", expr2);
+			dpGaussianParams.put("epsilon", expr3);
+			dpGaussianParams.put("delta", expr4);
+			currBuiltinOp = new ParameterizedBuiltinOp(target.getName(), DataType.MATRIX,
+				ValueType.FP64, ParamBuiltinOp.DP_GAUSSIAN, dpGaussianParams);
+			break;
+		}
 		case QUANTIZE_COMPRESS:
 			currBuiltinOp = new BinaryOp(target.getName(), target.getDataType(), target.getValueType(), OpOp2.valueOf(source.getOpCode().name()), expr, expr2);
 			break;

src/main/java/org/apache/sysds/runtime/controlprogram/context/ExecutionContext.java

@@ -62,6 +62,7 @@
 import org.apache.sysds.runtime.meta.MetaData;
 import org.apache.sysds.runtime.meta.MetaDataFormat;
 import org.apache.sysds.runtime.util.HDFSTool;
+import org.apache.sysds.runtime.privacy.dp.DPBudgetAccountant;
 import org.apache.sysds.utils.Statistics;
 
 import java.util.ArrayList;
@@ -90,6 +91,8 @@ public class ExecutionContext {
 
 	protected SEALClient _seal_client;
 
+	private DPBudgetAccountant _dpBudgetAccountant = null;
+
 	//parfor temporary functions (created by eval)
 	protected Set<String> _fnNames;
 
@@ -144,6 +147,12 @@ public void setLineage(Lineage lineage) {
 		_lineage = lineage;
 	}
 
+	public DPBudgetAccountant getDPBudgetAccountant() {
+		if (_dpBudgetAccountant == null)
+			_dpBudgetAccountant = new DPBudgetAccountant(1.0, 1e-5);
+		return _dpBudgetAccountant;
+	}
+
 	public boolean isAutoCreateVars() {
 		return _autoCreateVars;
 	}

src/main/java/org/apache/sysds/runtime/instructions/CPInstructionParser.java

@@ -65,6 +65,7 @@
 import org.apache.sysds.runtime.instructions.cp.UnaryCPInstruction;
 import org.apache.sysds.runtime.instructions.cp.VariableCPInstruction;
 import org.apache.sysds.runtime.instructions.cp.UnionCPInstruction;
+import org.apache.sysds.runtime.instructions.cp.DPBuiltinCPInstruction;
 import org.apache.sysds.runtime.instructions.cp.EinsumCPInstruction;
 import org.apache.sysds.runtime.instructions.cpfile.MatrixIndexingCPFileInstruction;
 
@@ -226,7 +227,10 @@ public static CPInstruction parseSingleInstruction ( InstructionType cptype, Str
 
 			case EINSUM:
 				return EinsumCPInstruction.parseInstruction(str);
-
+
+			case DPBuiltin:
+				return DPBuiltinCPInstruction.parseInstruction(str);
+
 			default:
 				throw new DMLRuntimeException("Invalid CP Instruction Type: " + cptype );
 		}

src/main/java/org/apache/sysds/runtime/instructions/cp/CPInstruction.java

@@ -48,7 +48,8 @@ public enum CPType {
 		EvictLineageCache, EINSUM,
 		NoOp,
 		Union,
-		QuantizeCompression
+		QuantizeCompression,
+		DPBuiltin
 	 }
 
 	protected final CPType _cptype;