diff --git a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/xml/StaxXmlGenerator.scala b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/xml/StaxXmlGenerator.scala
index 2a08a2e4d4b58..6e381a2974c79 100644
--- a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/xml/StaxXmlGenerator.scala
+++ b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/xml/StaxXmlGenerator.scala
@@ -19,11 +19,11 @@ package org.apache.spark.sql.catalyst.xml
 import java.io.Writer
 import java.sql.Timestamp
 import java.util.Base64
-import javax.xml.stream.XMLOutputFactory
 
 import scala.collection.Map
 
 import org.apache.hadoop.shaded.com.ctc.wstx.api.WstxOutputProperties
+import org.apache.hadoop.shaded.com.ctc.wstx.stax.WstxOutputFactory
 
 import org.apache.spark.SparkIllegalArgumentException
 import org.apache.spark.sql.catalyst.InternalRow
@@ -72,7 +72,13 @@ class StaxXmlGenerator(
   private val binaryFormatter = ToStringBase.getBinaryFormatter
 
   private val gen = {
-    val factory = XMLOutputFactory.newInstance()
+    // Instantiate the Woodstox factory directly from the shaded Hadoop classes instead of
+    // using XMLOutputFactory.newInstance(). The latter resolves an implementation via the
+    // service-loader mechanism, which could pick up a different (unshaded) StAX provider on the
+    // classpath. Such a provider would not understand the shaded WstxOutputProperties keys set
+    // below and would throw IllegalArgumentException. Constructing the shaded factory directly
+    // guarantees the properties and the implementation always match.
+    val factory = new WstxOutputFactory()
     // to_xml disables structure validation to allow multiple root tags
     factory.setProperty(WstxOutputProperties.P_OUTPUT_VALIDATE_STRUCTURE, validateStructure)
     factory.setProperty(WstxOutputProperties.P_OUTPUT_VALIDATE_NAMES, options.validateName)