TL;DR
This paper introduces a novel VM-based intrusion recovery architecture designed to minimize service downtime and data loss. By integrating techniques like cross-layer damage assessment and heterogeneous VM migration, their system aims to comprehensively clean infected systems while ensuring desired availability and continuity.
What Problem Does This Paper Address?
Intrusion recovery systems typically aim for comprehensive cleaning of infected systems. However, the complexity of modern systems, diverse attack vectors, and sophisticated attacker techniques make it challenging to completely "sweep the footprint" of an attack. Traditional recovery methods like system rebuilding or simply terminating and restarting attacked processes often lead to significant service downtime, loss of accumulated state, and failure to address secondary infection effects. This paper addresses the critical need for intrusion recovery that prioritizes availability and continuity, especially for mission-critical and business-critical applications that cannot tolerate extended downtime or data loss.
Methods
The authors propose a three-phase intrusion recovery framework:
- Routine Execution Phase: This phase focuses on non-intrusive checkpointing of system states and logging of system calls to track dependencies. This lays the groundwork for later damage assessment.
- Damage Assessment Phase: Upon detecting an intrusion, this phase aims to identify infected objects. It employs backward and forward taint tracking to pinpoint attack seeds and propagate the damage assessment. To maintain availability, the system can halt briefly at the start but then continue with taint-tracking-based quarantine.
- Recovery Phase: This phase deploys repairs based on the damage assessment. Infected objects are either removed or restored to a "good state." Coordinated repair is emphasized to ensure consistency. The system aims to achieve high availability by quarantining only infected processes while allowing uninfected ones to continue functioning.
Two systems are designed to realize this framework:
- SHELF (System Call Abstraction Level): Implemented at the VMM and host layers, SHELF uses state recording/restoration, logging/reconstruction, dynamic damage assessment, quarantine enforcement, and a recovery engine. It classifies objects into malicious, infected, and uninfected categories.
- PEDA (Extending SHELF to Defeat Damage Hiding Attacks): PEDA builds upon SHELF by incorporating heterogeneous VM migration and Revirt-style replay. It uses CPU-virtualization-based VMs for online operations and CPU-emulation-based VMs for offline, fine-grained damage assessment. This allows for comprehensive damage analysis without significantly impacting online service throughput.
Datasets
Not discussed in the paper. The authors evaluate their systems using simulated attacks and real-world malware scenarios (ARK rootkit and Lion Worm).
Key Insights and Ideas
- Availability is Paramount: The paper strongly argues that for many systems, intrusion recovery must prioritize availability and continuity, not just complete cleaning. This shifts the focus from a purely "clean system" goal to a "clean and running system" goal.
- VMs as a Foundation: Leveraging Virtual Machines (VMs) is key to isolating recovery components from the potentially compromised guest OS and enabling flexible migration and replay for damage assessment.
- Cross-Layer Damage Assessment: PEDA's approach of using heterogeneous VMs (virtualization for online, emulation for offline analysis) is crucial for achieving both high performance online and deep, fine-grained damage analysis offline.
- State Preservation: The emphasis on preserving accumulated state for stateful applications is vital for minimizing business continuity loss, a common drawback of simpler recovery methods.
- Damage Hiding Attacks: The motivation for PEDA highlights the evolving threat landscape, where attackers actively try to evade detection mechanisms, necessitating more sophisticated analysis techniques.
Key Results
- SHELF Damage Assessment: In experiments, SHELF successfully identified attack seeds and pinpointed infected objects for both the ARK rootkit and Lion Worm scenarios. For the ARK rootkit, damage assessment took 0.17 seconds for less than 500 objects. For a denser simulated graph of 1000 objects, it took 1.18 seconds.
- SHELF Availability Impact: During intensive recovery jobs, SHELF caused throughput losses of 21.4% for dbench and 12.4% for Apache ab, indicating a manageable impact on service quality.
- PEDA Availability Impact: PEDA aims to introduce less than 1% runtime overhead to online services due to its offline analysis approach.
Limitations and Future Work
- 24/7 Servers: The paper acknowledges the challenge of coping with servers that are constantly busy and have no downtime windows for offline repairs.
- Kernel Rootkits: Handling kernel rootkits and other deep kernel infections remains a challenging area for availability-sensitive intrusion recovery.
- Offline Repair Timing: PEDA's repair strategy relies on identifying low-traffic periods (e.g., 2-5 AM) for catch-up and repair, which might not be feasible for all systems.
- Future Work: The authors suggest further research into handling 24/7 servers and kernel-level infections. They also emphasize the need for community consideration of these open challenges.
Summary updated