Letting applications operate through attacks launched from compromised drivers

Shengzhi Zhang, Peng LiuAsiaCCS 2012 · DOI

TL;DR

This paper introduces DRASP, a novel system that allows applications to continue operating even when their underlying device drivers are compromised by attackers. By leveraging virtualized heterogeneous devices and running synchronized replicas, DRASP can detect driver-based attacks and seamlessly switch to a healthy replica, preserving application state and continuity.

What Problem Does This Paper Address?

In today's interconnected world, the availability and continuity of online services are paramount. While traditional security measures focus on protecting applications from their own vulnerabilities, a growing threat emerges from attackers exploiting vulnerabilities in device drivers. These drivers, which form a significant portion of operating system kernels, offer attackers privileged access to the entire system. Exploiting driver vulnerabilities allows attackers to manipulate applications, steal data, or disrupt services, often forcing a complete system reboot. This reboot leads to the loss of critical "work in progress" states, causing significant downtime and data loss. Existing recovery techniques are often insufficient against these driver-level attacks, necessitating a new approach to maintain service continuity.

Methods

DRASP tackles this challenge by employing a Diverse Replication strategy. The core idea is to run multiple replicas of the same service application, but crucially, these replicas are configured to use heterogeneous device drivers. This heterogeneity is achieved through virtualization, where different virtual hardware devices are emulated for each replica. The assumption is that drivers for different hardware (even if performing the same function) will have distinct vulnerabilities.

Here's how it works:

  • Heterogeneous Replicas: Two (or more) replicas of the service application are run, each within a virtual machine. These VMs are configured to emulate different hardware devices (e.g., different network interface cards, sound cards). Consequently, their guest operating systems load different device drivers.
  • Loose Synchronization: The replicas are synchronized at a system call level, ensuring that they process requests and maintain their states in a closely aligned manner. This synchronization is designed to be "loose" to avoid excessive overhead, but robust enough to detect discrepancies.
  • Proxy and Validation: A proxy sits in front of the replicas, splitting incoming requests to both. It then validates the responses and critical states of the two applications.
  • Intrusion Detection: If an attacker exploits a vulnerability in a driver on one replica, it will likely cause that replica's response or state to diverge from the other, healthy replica. The proxy detects this mismatch.
  • Operate-through Response: Upon detecting an intrusion, DRASP seamlessly transfers the workload to the surviving, uncompromised replica. This allows the service to continue operating without interruption, preserving its accumulated state and continuity. The compromised replica can then be safely restarted or replaced.

Datasets

Not discussed in the paper. The evaluation uses a simulated attack scenario on a web server (dhttpd) and a benchmarking tool (Apache benchmark).

Key Insights and Ideas

  • Driver Diversity as a Security Primitive: The paper highlights that exploiting driver vulnerabilities is a significant threat, and leveraging the inherent diversity of drivers (through virtualization) is a powerful defense mechanism.
  • The Power of Heterogeneity: By forcing replicas to use different drivers, an exploit targeting one driver is unlikely to affect the other, creating a natural isolation.
  • Seamless Failover for State Preservation: DRASP's ability to detect and immediately switch to a healthy replica is crucial for maintaining service continuity and preventing data loss, even during driver-level attacks.
  • Virtualization as an Enabler: The use of virtual platforms is key to easily creating the necessary hardware and driver diversity without requiring physical hardware changes.

Key Results

  • Intrusion Detection: The paper demonstrates successful detection of a simulated buffer overflow attack on a network driver. The response validator detected the mismatch between the compromised and healthy replicas.
  • Efficiency: The overhead introduced by DRASP (synchronization, validation) was found to be minimal. For a web server, the response time increased by only a few milliseconds (from 45ms to 47ms for coarse-grained, and 50ms for fine-grained validation).
  • Continuity and State Preservation: Experiments showed that DRASP could maintain service continuity and preserve accumulated state without any missing or false responses during simulated attacks and replica switching.

Limitations and Future Work

  • Encrypted Traffic: The current response validation mechanism struggles with encrypted traffic (like HTTPS) because the back-stage replica cannot decrypt and understand the requests, making response comparison meaningless. Future work aims to address this.
  • Resource Overhead: While the overhead is shown to be reasonable, running multiple replicas inherently requires more computational resources.
  • Configuration Complexity: Developers need to identify critical code/data sections for state validation, which requires domain knowledge.
  • Future Work: The authors suggest extending the approach to handle encrypted traffic and exploring more sophisticated synchronization and validation techniques. They also note that DRASP is compatible with existing fault-tolerant systems.

Summary updated

BibTeX
@inproceedings{zhang2012letting,
  author = {Shengzhi Zhang and Peng Liu},
  title = {{Letting applications operate through attacks launched from compromised drivers}},
  booktitle = {AsiaCCS},
  year = {2012},
  doi = {10.1145/2414456.2414510},
}
Figure