Fix heap OOB read in SVMRegressor via attribute size validation

onnxruntime/core/providers/cpu/ml/svmregressor.cc

@@ -28,8 +28,25 @@ SVMRegressor<T>::SVMRegressor(const OpKernelInfo& info)
   auto onec = info.GetAttrOrDefault<int64_t>("one_class", 0);
   one_class_ = (onec != 0);
 
+  ORT_ENFORCE(!rho_.empty(), "SVMRegressor: rho must not be empty");
+
   if (vector_count_ > 0) {
+    // Validate attribute array sizes against declared dimensions to prevent
+    // out-of-bounds reads from crafted models.
+    ORT_ENFORCE(coefficients_.size() >= static_cast<size_t>(vector_count_),
+                "SVMRegressor: coefficients size (", coefficients_.size(),
+                ") must be >= n_supports (", vector_count_, ")");
+    ORT_ENFORCE(!support_vectors_.empty(),
+                "SVMRegressor: support_vectors must not be empty when n_supports > 0");
+    ORT_ENFORCE(support_vectors_.size() % static_cast<size_t>(vector_count_) == 0,
+                "SVMRegressor: support_vectors size (", support_vectors_.size(),
+                ") must be a multiple of n_supports (", vector_count_, ")");
+
     feature_count_ = support_vectors_.size() / vector_count_;  // length of each support vector
+    ORT_ENFORCE(feature_count_ > 0,
+                "SVMRegressor: support_vectors size (", support_vectors_.size(),
+                ") is too small for n_supports (", vector_count_, ")");
+
     mode_ = SVM_TYPE::SVM_SVC;
   } else {
     feature_count_ = coefficients_.size();

onnxruntime/test/providers/cpu/ml/svmregressor_test.cc

@@ -136,5 +136,49 @@ TEST(MLOpTest, SVMRegressorLinear) {
   test.Run();
 }
 
+TEST(MLOpTest, SVMRegressorUndersizedCoefficients) {
+  OpTester test("SVMRegressor", 1, onnxruntime::kMLDomain);
+
+  std::vector<float> coefficients = {1.f};  // needs 5, only 1 provided
+  std::vector<float> support_vectors = {0.f, 0.5f, 32.f, 1.f, 1.5f, 1.f, 2.f, 2.9f, -32.f,
+                                        12.f, 12.9f, -312.f, 43.f, 413.3f, -114.f};
+  std::vector<float> rho = {0.1f};
+  std::vector<float> kernel_params = {0.001f, 0.f, 3.f};
+
+  test.AddAttribute("kernel_type", std::string("RBF"));
+  test.AddAttribute("coefficients", coefficients);
+  test.AddAttribute("support_vectors", support_vectors);
+  test.AddAttribute("rho", rho);
+  test.AddAttribute("kernel_params", kernel_params);
+  test.AddAttribute("n_supports", static_cast<int64_t>(5));
+
+  test.AddInput<float>("X", {1, 3}, {1.f, 0.f, 0.4f});
+  test.AddOutput<float>("Y", {1, 1}, {0.f});
+
+  test.Run(OpTester::ExpectResult::kExpectFailure, "coefficients size");
+}
+
+TEST(MLOpTest, SVMRegressorUndersizedSupportVectors) {
+  OpTester test("SVMRegressor", 1, onnxruntime::kMLDomain);
+
+  std::vector<float> coefficients = {1.f, 1.f, 1.f, 1.f, 1.f};
+  std::vector<float> support_vectors = {0.1f, 0.2f};  // far too small for n_supports=5
+  std::vector<float> rho = {0.1f};
+  std::vector<float> kernel_params = {0.001f, 0.f, 3.f};
+
+  test.AddAttribute("kernel_type", std::string("RBF"));
+  test.AddAttribute("coefficients", coefficients);
+  test.AddAttribute("support_vectors", support_vectors);
+  test.AddAttribute("rho", rho);
+  test.AddAttribute("kernel_params", kernel_params);
+  test.AddAttribute("n_supports", static_cast<int64_t>(5));
+
+  test.AddInput<float>("X", {1, 3}, {1.f, 0.f, 0.4f});
+  test.AddOutput<float>("Y", {1, 1}, {0.f});
+
+  // feature_count_ = 2/5 = 0, which triggers the feature_count_ > 0 check
+  test.Run(OpTester::ExpectResult::kExpectFailure, "support_vectors size");
+}
+
 }  // namespace test
 }  // namespace onnxruntime