Abstract
Distributed data analytics platforms like Apache Spark are designed for high-performance scale, often at the expense of fine-grained security. Existing security solutions, typically based on Inline Reference Monitors (IRM), are vulnerable to "transient evasion attacks"—stealthy exploits where users abuse internal platform APIs to bypass security controls. This paper introduces SecureDL, a defense-in-depth framework that combines proactive static analysis with reactive runtime sandboxing. SecureDL screens user code for suspicious API usage before execution and enforces runtime checks via a hardened Java Security Manager and binary integrity verification. Evaluation on TPC-H and HiBench benchmarks demonstrates that SecureDL effectively blocks evasion attacks with a modest performance overhead of approximately 4%.
Introduction
As distributed analytics platforms become central to enterprise data processing, the need for robust access control has intensified. However, the authors identify a critical flaw in current security models: transient evasion attacks. In these scenarios, a malicious insider uses legitimate platform APIs (such as Java Reflection or internal package calls) to access raw data before security filters are applied. Because these operations occur entirely within the memory of the running job, they leave no traces in system logs. Standard sandboxing techniques, such as the default Java Security Manager, are often insufficient to prevent these sophisticated workarounds.
The SecureDL Framework
SecureDL addresses these vulnerabilities through a two-layered defense strategy that integrates into the existing Spark ecosystem.
Proactive Defense
Before a job is executed, SecureDL performs a "pre-flight check" using static program analysis. It screens user-submitted code for blockable signatures, such as attempts to declare classes within protected internal packages (e.g., org.apache.spark). Additionally, it employs "soundy" dataflow analysis to detect suspicious but non-blockable API usage, such as using reflection to traverse an RDD’s internal pointers to access unauthorized data.
Reactive Defense
To catch threats that might bypass static analysis, SecureDL employs a runtime reactive layer. This includes binary integrity checking to ensure the Trusted Computing Base (TCB) has not been tampered with, and a hardened Java Security Manager to block unauthorized system calls. Furthermore, runtime instrumentation injects security checks that trigger exceptions if code attempts to invoke sensitive APIs on protected data objects during execution.
Experimental Evaluation
The authors evaluated SecureDL using industry-standard benchmarks and real-world data:
- Performance Overhead: Testing on a 6-node Hadoop cluster using TPC-H queries showed an average overhead of only 4%. While complex queries involving multiple joins saw higher overhead (up to 16%), simpler aggregation tasks remained highly efficient.
- Security Effectiveness: In a test suite of 15 custom-crafted transient evasion attacks, the proactive analysis successfully detected and blocked every attempt.
- Real-world Compatibility: The authors analyzed 2,120 real-world Spark repositories from GitHub. Of the analyzable jars, the system exhibited a very low false-positive rate, only flagging highly unusual developer behaviors such as manual external network connections or internal package redefinitions.
Conclusion
SecureDL provides a pragmatic and effective solution for securing distributed analytics platforms against sophisticated API-based attacks. By combining the gatekeeping capabilities of static analysis with the bodyguard-like protection of runtime sandboxing, it achieves high security without sacrificing the performance that makes platforms like Spark valuable. Future work aims to extend these protections to non-JVM interfaces such as PySpark and R.
Summary updated