Introduction
Modern computing systems inadvertently leak sensitive information through physical side channels, such as power consumption, temperature, and electromagnetic emissions. These signals often correlate strongly with processor activity, allowing attackers to employ advanced signal processing and machine learning (ML) to infer private data. For instance, an attacker might identify a visited website, capture keystrokes, or extract encryption keys simply by monitoring power fluctuations.
Traditional defenses, such as maintaining constant power consumption or injecting random noise, face significant hurdles. Achieving perfectly flat power profiles is technically challenging on commodity hardware due to unavoidable transition leaks. Furthermore, simple noise injection is often insufficient against sophisticated ML models that can filter out non-correlated signals. Maya addresses these vulnerabilities by introducing a software-based defense that "falsifies" the power signal rather than merely suppressing it.
System Architecture
Maya is designed as an application-transparent defense layer, implementable within the operating system or as a hardware component. Its primary goal is to reshape the system's power consumption into a deceptive "fake" profile that mimics legitimate application behavior.
The system consists of two core components:
- Mask Generator: This component defines the target "fake" power profile. Instead of a static target, it generates a dynamic signal—specifically a Gaussian Sinusoid—that incorporates both periodic oscillations and stochastic noise. This complexity is intended to simulate the natural phase behavior and frequency characteristics of real software.
- MIMO Controller: Utilizing Multiple Input Multiple Output (MIMO) robust control theory, the controller manages three distinct system inputs to track the target power profile:
- DVFS (Dynamic Voltage and Frequency Scaling): Adjusts the CPU clock speed.
- Idle Threads: Reduces power by injecting periods of inactivity.
- Balloon Threads: Increases power by executing dummy floating-point operations.
By sampling power data via interfaces like Intel's RAPL and adjusting these "knobs" every 20ms, the controller ensures the actual power consumption follows the deceptive target with high fidelity.
The Falsification Strategy
A key insight of the Maya project is the shift from signal suppression to signal falsification. While suppression attempts to hide information (often unsuccessfully), falsification provides the attacker with "misinformation." By making the power trace look like a different, valid application, Maya effectively camouflages the victim application's real signature.
The choice of a Gaussian Sinusoid mask is critical. A flat power mask is easily identified as artificial by an adaptive attacker. In contrast, the Gaussian Sinusoid introduces artificial execution phases and frequency peaks that resemble the loops and branches of a real program. This makes it significantly harder for an attacker to distinguish the mask from the underlying computation.
Evaluation and Performance
The researchers evaluated Maya against state-of-the-art ML classifiers, specifically Multi-Layer Perceptrons (MLPs), using standard benchmarks like PARSEC and SPLASH2x. The results demonstrate a dramatic reduction in attack efficacy:
- Application Identification: On unprotected systems, attackers could identify running applications with 92% accuracy. With Maya, this dropped to 20%, nearing the random chance threshold of 13%.
- Data Privacy: In tests involving video transcoding (
ffmpeg), attacker accuracy in identifying specific videos fell from 99% to 39% (random chance being 33%).
Despite its effectiveness, Maya introduces a notable performance trade-off. The system typically causes execution slowdowns between 50% and 100%. Interestingly, it also resulted in a 29% reduction in average power consumption compared to a high-performance baseline, as the controller often throttles the system to meet the mask's targets.
Conclusion
Maya represents a significant step forward in side-channel defense by leveraging control theory to provide robust power falsification. While the current software implementation is effective against millisecond-scale attacks (like application or keystroke detection), its performance overhead remains a challenge for general-purpose use. Future work may involve hardware-level implementations to reduce overhead and address microsecond-scale threats, such as those targeting cryptographic operations.
Summary updated