Abstract
ShadowScope is a novel monitoring framework designed to secure GPUs by leveraging performance counters as side-channel signals. By decomposing kernel execution into modular, "composable" segments, ShadowScope overcomes the inherent scheduling noise and variability of parallel hardware. This approach enables the detection of diverse attacks, including buffer overflows and Rowhammer, with high accuracy (96% average TPR) and minimal performance overhead (4.6% in hardware-assisted mode).
Introduction
As GPUs become the backbone of modern AI and data centers, they have also become prime targets for attackers. Vulnerabilities range from traditional memory-safety issues, such as buffer overflows, to sophisticated microarchitectural threats like Rowhammer. The standard defense—"golden model" validation—involves comparing a kernel’s runtime behavior against a trusted reference trace. However, this is notoriously difficult on GPUs because they are massively parallel and use dynamic scheduling. Two "identical" runs can produce wildly different timing and signal traces, making traditional monolithic golden models fragile and prone to high false-alarm rates.
Methodology
ShadowScope introduces two major innovations to make GPU validation practical:
- Composable Golden Models: Instead of matching a single complex trace, ShadowScope decomposes kernel execution into modular, repeatable segments. It uses lightweight "composable functions"—markers placed at kernel boundaries or internal phases—to segment the data. This allows the system to validate each piece of the execution independently, making it resilient to scheduling jitter.
- Covert Channel Markers: To align these segments without heavy performance penalties, specific atomic operations (like
atomicCAS) are used to trigger distinct, recognizable spikes in performance counters, acting as "sync pulses" for the validator.
The framework was evaluated in two versions: a software implementation using NVIDIA’s CUPTI profiling API and a proposed hardware-assisted version (ShadowScope+) featuring a local Performance Monitoring Unit (PMU) and an on-chip Validator Module for real-time integrity checks.
Evaluation and Results
The researchers evaluated ShadowScope against 18 benchmarks (Rodinia, CUDA SDK, GraphBig), four DNN architectures (CifarNet, AlexNet, SqueezeNet, ResNet-50), and four representative attack scenarios: buffer overflows, mind-control attacks (skipping DNN layers), Rowhammer, and DoS/slowdown attacks.
Performance and Accuracy
- High Detection Accuracy: The software implementation achieved a 96% average True Positive Rate (TPR) for buffer overflows and a 100% TPR for Rowhammer and DoS attacks.
- Low False Positives: The system achieved false positive rates as low as 0% under controlled conditions, though rates slightly increased for extremely fast-executing kernels where sampling resolution was limited.
- Minimal Overhead: The hardware-assisted ShadowScope+ incurred an average runtime overhead of just 4.6%, making it practical for "always-on" security.
Conclusion
ShadowScope demonstrates that GPU performance counters can be effectively repurposed as security side-channels. By validating execution at a functional, modular level, the framework successfully "absorbs" scheduling noise that typically plagues GPU monitoring. While the software version is currently limited by sampling rates, the simulated hardware-assisted version offers a path toward low-overhead, real-time GPU security. Future work includes exploring adaptive thresholds and extending the framework to multi-GPU configurations.
Summary updated