TL;DR
This paper introduces Timing Information Flow Control (TIFC), a novel approach to manage sensitive information leakage through timing channels in cloud environments. By extending existing information flow control techniques with "timing labels," TIFC aims to provide a framework for building more secure cloud infrastructure that can still leverage statistical multiplexing for efficiency.
What Problem Does This Paper Address?
Cloud computing's reliance on massive parallelism and shared resources creates a fertile ground for timing side-channel attacks. These attacks exploit subtle variations in the execution time of computations to infer sensitive information, bypassing traditional security mechanisms that focus on data content. While techniques exist to protect data itself, they offer little defense against these timing leaks. The paper highlights four key issues exacerbating this problem in the cloud: the pervasiveness of timing channels due to parallelism, the ease of mounting attacks (requiring only purchased CPU time), the difficulty of detecting such attacks, and the conflict between traditional timing channel mitigation (resource reservation) and the cloud's business model of statistical multiplexing.
Methods
The core contribution is Timing Information Flow Control (TIFC), an extension of decentralized information flow control (DIFC) techniques. TIFC introduces timing labels alongside content labels for system objects (processes, files, messages). These timing labels specify the maximum rate (in bits per second) at which information from a particular user might leak via timing events associated with that object.
Key components of the TIFC approach include:
- Timing Labels: Represented as
{Lc/LT}, whereLcis the set of content tags andLTis the set of timing tags. Timing tags are of the formUf, indicating userU's information leaking at ratef. An infinite rate is denoted by∞. - Declassification Capabilities: Similar to content declassification, TIFC adds timing declassification capabilities (
UForUf) that allow a process to explicitly reduce the timing taint of information it handles. - System Design Tools:
- Deterministic Execution: This technique, inspired by systems like Determinator, aims to decouple a process's execution timing from its content. By preventing processes from inferring time through explicit inputs, it makes it safer for content and timing labels to differ.
- Pacing Queues: These queues limit the rate at which messages are released, effectively "downgrading" timing tags to a controlled frequency. This allows for statistical multiplexing while ensuring timing leaks are bounded.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Timing is a First-Class Security Concern: The paper elevates timing information from an implementation detail to a first-class security concern, deserving explicit labeling and control mechanisms.
- Bridging the Gap Between Security and Cloud Efficiency: TIFC offers a potential solution to the fundamental tension between robust timing channel security and the cloud's need for statistical multiplexing and resource utilization.
- Labeling the "How" Not Just the "What": Unlike traditional information flow control, TIFC labels not only the content of data but also the manner in which it might be leaked through timing.
- Practicality over Elimination: The paper acknowledges that completely eliminating timing channels is likely infeasible. Instead, TIFC focuses on controlling and limiting the rate of leakage.
Key Results
The paper presents case studies illustrating how TIFC can be applied in different cloud scenarios:
- Private Per-Client Hardware Resources: Demonstrates how TIFC labels can enforce isolation when dedicated hardware is used, ensuring Alice's job completion time is independent of Bob's.
- Shared Resource with Reservation-based Scheduling: Shows how TIFC, combined with a "demand-insensitive" reservation scheduler, can maintain timing isolation even when resources are shared, albeit with potential underutilization.
- Shared Resource with Demand-driven Scheduling: This scenario, which allows for statistical multiplexing, shows how TIFC, coupled with pacing queues, can limit timing leaks to an agreed-upon rate
f, even when job completion times are influenced by other users. For example, Alice's job results can be released at ratef, leaking at mostfbits per second of Bob's information.
Limitations and Future Work
The authors explicitly state that this paper "merely explores one potential approach, which has not been rigorously formalized or experimentally validated."
Future work includes:
- Formalization: Rigorously formalizing the TIFC model and its security arguments.
- Experimental Validation: Implementing TIFC in a system like Determinator [4] to test its effectiveness and performance in real-world scenarios.
- Extending to Other Channels: While the current focus is on internal cloud timing channels, the authors suggest the model may extend to other channels, such as network communication.
Summary updated