Adding Random Forest EzPC code

EzPC/EzPC/test_suite/random_forest.ezpc

+(* Random Forest implementation for inference on California Housing Dataset -- Regression *)
+
+(* These values come from a python program that I used to train
+the Random Forest. Accuracy on testing set (20% of full dataset) = 83.04 *)	
+uint32 noOfFeatures = 13;
+(* This depth is the depth in python code + 1 *)
+uint32 maxDepth = 11;
+int32 maxDepthInt = 11;
+uint32 noOfTrees = 10;
+uint32 noOfNodes = 3035; (* (2^maxDepth) *)
+
+(* To get secret shares of an element from array when the element's index is secret shared *)
+def void accessElementOneD(uint64_bl[noOfFeatures] arr, uint64_bl idx, uint64_bl[1] result)
+{
+	result[0] = 0uL;
+	uint64_pl ctr = 0uL;
+	for i=[0:noOfFeatures]
+	{
+		result[0] = result[0] +_bl (idx == ctr?arr[i]:0uL);
+		ctr = ctr + 1uL;	
+	};
+}
+
+def void accessElementTwoD(uint64_bl[noOfTrees][noOfNodes] arr, int64_pl treeId, uint64_al idx, uint64_bl[1] result, uint32_pl start, uint32_pl end)
+{
+	result[0] = 0uL;
+	uint64_pl ctr = 0uL;
+	uint64_bl idxcpy = idx;
+	for i=[start:end]
+	{
+		result[0] = result[0] +_bl (idxcpy == ctr?arr[treeId][i]:0uL);
+		ctr = ctr + 1uL;	
+	};
+}
+
+def uint32_pl pow(uint32_pl base, int32_pl power)
+{
+	uint32_pl result = 1;
+	for i=[0:power]
+	{
+		result = result * base;
+	};
+	return result;
+}
+
+def void infer(uint64_bl[noOfTrees][noOfNodes] modelFeatureChoice, 
+		uint64_bl[noOfTrees][noOfNodes] modelThresholdValue,
+		uint64_bl[noOfFeatures] query,
+		uint64_bl[noOfTrees] result)
+{
+	uint64_al currentIndex = 0uL;
+	uint64_al leftOrRight = 0uL;
+	uint64_bl[1] currentSharedFeatureChoice;
+	uint64_bl[1] currentSharedThreshold;
+	uint64_bl[1] currentSharedQueryValue;
+	currentSharedFeatureChoice[0] = 0uL;
+	currentSharedThreshold[0] = 0uL;
+	currentSharedQueryValue[0] = 0uL;
+	uint32_pl startIdx = 0;
+	uint32_pl endIdx = 0;
+
+	for i=[0:noOfTrees]
+	{
+		currentIndex = 0uL;
+		
+		(* Root level *)
+		currentSharedFeatureChoice[0] = modelFeatureChoice[i][0];
+		currentSharedThreshold[0] = modelThresholdValue[i][0];
+		accessElementOneD(query, currentSharedFeatureChoice[0], currentSharedQueryValue);
+		leftOrRight = (currentSharedThreshold[0] > currentSharedQueryValue[0])?1L:2L;
+		
+		for j=[1:maxDepthInt-1]
+		{
+			currentIndex = (currentIndex *_al 2uL) +_al leftOrRight;				
+			startIdx = pow(2u, j)-1u;
+			endIdx = pow(2u, j+1)-1u;
+			accessElementTwoD(modelThresholdValue, i, currentIndex, currentSharedThreshold, startIdx, endIdx);
+			accessElementTwoD(modelFeatureChoice, i, currentIndex, currentSharedFeatureChoice, startIdx, endIdx);
+			accessElementOneD(query, currentSharedFeatureChoice[0], currentSharedQueryValue);
+			leftOrRight = (currentSharedThreshold[0] > currentSharedQueryValue[0])?1L:2L;
+		};
+		
+		(* Leaf level *)
+		currentIndex = (currentIndex *_al 2uL) +_al leftOrRight;				
+		startIdx = pow(2u, maxDepthInt-1)-1u;
+		endIdx = pow(2u, maxDepthInt)-1u;
+		accessElementTwoD(modelThresholdValue, i, currentIndex, currentSharedThreshold, startIdx, endIdx);
+		result[i] = currentSharedThreshold[0];
+	};
+}
+
+def void main()
+{
+	(* Taking inputs as boolean shares because comparison are performed *)
+	uint64_bl[noOfFeatures] inferenceQuery;
+	uint64_bl[noOfTrees][noOfNodes] modelFeatureChoice;
+	(* Threshold value for leaves = inference output from that tree *)
+	uint64_bl[noOfTrees][noOfNodes] modelThresholdValue;
+	uint64_bl[noOfTrees] inferenceResult;
+	uint64_bl finalResult = 0uL;	
+
+	(* One party inputs the inference query *)
+	for i=[0:noOfFeatures]
+	{
+		inferenceQuery[i] = 20uL;	
+	};
+
+	(* Other party inputs the trained model in the form of the feature choice
+	at each node as well as the threshold value of the feature at that node *)
+	for i=[0:noOfTrees]
+	{
+		for j=[0:noOfNodes]
+		{
+			modelFeatureChoice[i][j] = (((noOfNodes))%noOfFeatures)+0uL;
+			modelThresholdValue[i][j] = (noOfNodes)+351uL;
+		};
+	};
+
+	infer(modelFeatureChoice, modelThresholdValue, inferenceQuery, inferenceResult);	
+	
+	(* Take mean of all the values in inference result *)
+	for i=[0:noOfTrees]
+	{
+		finalResult = finalResult +_al inferenceResult[i];
+	};
+	(* finalResult = finalResult/(noOfTrees+0uL); *)
+	output(ALL, finalResult)
+}

EzPC/EzPC/test_suite/random_forest_polish.ezpc

+(* Random Forest implementation for inference on Polish companies backruptcy Dataset (on UCI repo) -- Classification *)
+
+(* Accuracy numbers from python code -- imputation used is Mean because dataset has empty cells at places:
+This dataset is divided into 5 parts: 1year, 2 year, 3 year, 4 year and 5 year
+
+Depth = 12
+Trees = 10
+
+1 year: 95.1
+2 year: 91.6
+3 year: 89.7
+4 year: 88.7
+5 year: 92.4 
+
+Depth = 11
+Trees = 5
+
+1 year: 92.1
+2 year: 87.3
+3 year: 86.7
+4 year: 84.9
+5 year: 89.0
+
+*)
+
+
+
+uint32 noOfFeatures = 64;
+(* This depth is the depth in python code + 1 *)
+uint32 maxDepth = 11;
+int32 maxDepthInt = 11;
+uint32 noOfTrees = 5;
+uint32 noOfNodes = 2048; (* (2^maxDepth) *)
+
+(* To get secret shares of an element from array when the element's index is secret shared *)
+def void accessElementOneD(uint64_bl[noOfFeatures] arr, uint64_bl idx, uint64_bl[1] result)
+{
+	result[0] = 0uL;
+	uint64_pl ctr = 0uL;
+	for i=[0:noOfFeatures]
+	{
+		result[0] = result[0] +_bl (idx == ctr?arr[i]:0uL);
+		ctr = ctr + 1uL;	
+	};
+}
+
+def void accessElementTwoD(uint64_bl[noOfTrees][noOfNodes] arr, int64_pl treeId, uint64_al idx, uint64_bl[1] result, uint32_pl start, uint32_pl end)
+{
+	result[0] = 0uL;
+	uint64_pl ctr = 0uL;
+	uint64_bl idxcpy = idx;
+	for i=[start:end]
+	{
+		result[0] = result[0] +_bl (idxcpy == ctr?arr[treeId][i]:0uL);
+		ctr = ctr + 1uL;	
+	};
+}
+
+def uint32_pl pow(uint32_pl base, int32_pl power)
+{
+	uint32_pl result = 1;
+	for i=[0:power]
+	{
+		result = result * base;
+	};
+	return result;
+}
+
+def void infer(uint64_bl[noOfTrees][noOfNodes] modelFeatureChoice, 
+		uint64_bl[noOfTrees][noOfNodes] modelThresholdValue,
+		uint64_bl[noOfFeatures] query,
+		uint64_bl[noOfTrees] result)
+{
+	uint64_al currentIndex = 0uL;
+	uint64_al leftOrRight = 0uL;
+	uint64_bl[1] currentSharedFeatureChoice;
+	uint64_bl[1] currentSharedThreshold;
+	uint64_bl[1] currentSharedQueryValue;
+	currentSharedFeatureChoice[0] = 0uL;
+	currentSharedThreshold[0] = 0uL;
+	currentSharedQueryValue[0] = 0uL;
+	uint32_pl startIdx = 0;
+	uint32_pl endIdx = 0;
+
+	for i=[0:noOfTrees]
+	{
+		currentIndex = 0uL;
+		
+		(* Root level *)
+		currentSharedFeatureChoice[0] = modelFeatureChoice[i][0];
+		currentSharedThreshold[0] = modelThresholdValue[i][0];
+		accessElementOneD(query, currentSharedFeatureChoice[0], currentSharedQueryValue);
+		leftOrRight = (currentSharedThreshold[0] > currentSharedQueryValue[0])?1L:2L;
+		
+		for j=[1:maxDepthInt-1]
+		{
+			currentIndex = (currentIndex *_al 2uL) +_al leftOrRight;				
+			startIdx = pow(2u, j)-1u;
+			endIdx = pow(2u, j+1)-1u;
+			accessElementTwoD(modelThresholdValue, i, currentIndex, currentSharedThreshold, startIdx, endIdx);
+			accessElementTwoD(modelFeatureChoice, i, currentIndex, currentSharedFeatureChoice, startIdx, endIdx);
+			accessElementOneD(query, currentSharedFeatureChoice[0], currentSharedQueryValue);
+			leftOrRight = (currentSharedThreshold[0] > currentSharedQueryValue[0])?1L:2L;
+		};
+		
+		(* Leaf level *)
+		currentIndex = (currentIndex *_al 2uL) +_al leftOrRight;				
+		startIdx = pow(2u, maxDepthInt-1)-1u;
+		endIdx = pow(2u, maxDepthInt)-1u;
+		accessElementTwoD(modelThresholdValue, i, currentIndex, currentSharedThreshold, startIdx, endIdx);
+		result[i] = currentSharedThreshold[0];
+	};
+}
+
+def void main()
+{
+	(* Taking inputs as boolean shares because comparison are performed *)
+	uint64_bl[noOfFeatures] inferenceQuery;
+	uint64_bl[noOfTrees][noOfNodes] modelFeatureChoice;
+	(* Threshold value for leaves = inference output from that tree *)
+	uint64_bl[noOfTrees][noOfNodes] modelThresholdValue;
+	uint64_bl[noOfTrees] inferenceResult;
+	uint64_bl finalResult = 0uL;	
+
+	(* One party inputs the inference query *)
+	for i=[0:noOfFeatures]
+	{
+		inferenceQuery[i] = 20uL;	
+	};
+
+	(* Other party inputs the trained model in the form of the feature choice
+	at each node as well as the threshold value of the feature at that node *)
+	for i=[0:noOfTrees]
+	{
+		for j=[0:noOfNodes]
+		{
+			modelFeatureChoice[i][j] = (((noOfNodes))%noOfFeatures)+0uL;
+			modelThresholdValue[i][j] = (noOfNodes)+351uL;
+		};
+	};
+
+	infer(modelFeatureChoice, modelThresholdValue, inferenceQuery, inferenceResult);	
+	
+	(* Take mean of all the values in inference result *)
+	for i=[0:noOfTrees]
+	{
+		finalResult = finalResult +_al inferenceResult[i];
+	};
+	(* finalResult = finalResult/(noOfTrees+0uL); *)
+	output(ALL, finalResult)
+}