Automated Attack Discovery in Data Plane Systems

Qiao Kang, Jiarong Xing, Ang ChenCSET @ USENIX Security Symposium 2019 · OpenReview

TL;DR

This paper introduces "sensitivity attacks," a new class of vulnerabilities in programmable data plane systems. The authors propose an automated framework to discover these attacks by analyzing program behavior and then synthesize monitors to detect and defend against them. This work is crucial for securing the increasingly dynamic and programmable network infrastructure.

What Problem Does This Paper Address?

Modern network switches are becoming programmable, allowing for custom protocols and sophisticated packet processing directly in hardware. While this offers immense flexibility and performance, it also opens new security avenues. Traditional security often relies on control plane intelligence, but programmable data planes can react to network traffic much faster, making them vulnerable to manipulation. The authors identify a gap in understanding and addressing security risks specific to these programmable data planes, particularly how an attacker could exploit their sensitivity to input traffic patterns to disrupt normal operations. They aim to automate the discovery of these "sensitivity attacks" and the generation of defenses.

Methods

The paper proposes a three-step automated approach:

  1. Establishing Expected Behaviors: They use probabilistic symbolic execution to analyze a given P4 data plane program. This technique enumerates all possible execution paths and annotates them with probabilities, effectively defining the system's "normal" or "expected" behavior under various traffic conditions. They group these behaviors into equivalence classes (ECs) based on how packets are processed.
  2. Negating Expected Behaviors (Attack Discovery): The core of attack discovery involves generating malicious traffic patterns that deviate from the expected behavior. The authors propose an algorithm that greedily selects target ECs and uses directed symbolic execution to find packet sequences that maximize the deviation in probability distributions across ECs. This quantifies the "attack strength." They also address challenges like handling stateful behaviors and hash tables by modeling them as "greyboxes" with pre-determined collision rates.
  3. Synthesizing Runtime Monitors: Once malicious patterns are identified, the system synthesizes P4 monitors. These monitors are integrated into the original data plane program and use counters to track traffic distribution across ECs. A statistical check periodically compares current traffic patterns against expected ones, raising an alarm if significant deviations are detected. They also explore techniques like Count-Min sketches for memory-efficient monitoring.

Datasets

Not discussed in the paper. The authors use a P4 program for a load balancer as an example and conduct experiments in a Mininet environment.

Key Insights and Ideas

  • Programmability Creates New Attack Surfaces: The very dynamism and sensitivity that make programmable data planes powerful also make them susceptible to "sensitivity attacks" where crafted traffic can flip expected behaviors.
  • Automated Discovery is Key: Manually identifying all potential attack vectors in complex, dynamic data plane programs is infeasible. An automated approach is necessary for effective security.
  • Probabilistic Symbolic Execution for Behavior Analysis: This technique is crucial for understanding not just what a data plane program does, but how likely it is to do it under different traffic conditions, forming the basis for defining "normal."
  • The Tension Between Dynamicity and Security: There's an inherent trade-off: highly dynamic systems that react quickly to network events are harder to secure with traditional authentication mechanisms, which often require slower control plane intervention.
  • Greybox Modeling for Complex Components: Treating complex hardware components like hash tables as "greyboxes" with probabilistic collision rates allows for symbolic execution without needing full knowledge of their internal workings, making the analysis scalable.

Key Results

The paper presents a preliminary experimental validation using a simple P4 load balancer.

  • Attack Discovery: The system successfully identified two equivalence classes (ECs) for the load balancer, corresponding to packets being sent to port 0 or port 1 based on the sport%2 hash. It then generated attack patterns that could force the traffic distribution to be 100% on one port and 0% on the other, deviating from the expected 50/50 split.
  • Attack Feasibility: An attacker injecting a crafted traffic trace was able to double the traffic rate on one port while leaving the other under-utilized, demonstrating a successful degradation-of-service attack.
  • Monitoring Effectiveness: A manually written monitor successfully detected these skewed traffic distributions and raised an alarm, validating the monitoring approach.

Limitations and Future Work

  • Scalability of Symbolic Execution: While probabilistic symbolic execution is powerful, the paper acknowledges that state explosion can still be an issue, especially for complex programs or long sequences of packets.
  • Monitor Compression: Optimizing monitor resource usage (memory) is a significant challenge for complex data plane systems, and the paper mentions using sketching as a tentative approach.
  • Validation on Diverse Systems: The authors state that their current experiments are preliminary and that they are working on validating their methodology on a more diverse set of data plane systems.
  • Automated Monitor Synthesis: While the paper proposes synthesizing monitors, the current experimental setup used a manually written monitor. Automating this synthesis is a key part of their ongoing work.

Summary updated

BibTeX
@inproceedings{kang2019automated,
  author = {Qiao Kang and Jiarong Xing and Ang Chen},
  title = {{Automated Attack Discovery in Data Plane Systems}},
  booktitle = {CSET @ USENIX Security Symposium},
  year = {2019},
  url = {https://openreview.net/forum?id=RTeclpdu65},
}
Figure