HAWC-Leak: Heterogeneous Architecture Workload Classification utilizing Side-Channel Leakage

summary

Video file (mp4)

The gist

The gist: HAWC-Leak presents a dynamic workload classification framework that identifies specific hardware accelerator tasks in real time using non-invasive power analysis and achieves 100%

In short

HAWC-Leak uses non-invasive power analysis to classify hardware accelerator tasks in real time within complex, mixed-activity systems. The method employs a two-stage hierarchical Random Forest classifier to achieve 100% accuracy across AES, ECC, SHA-256, CNN, and idle states. This demonstrates that power consumption acts as a high-fidelity barcode for heterogeneous hardware execution.

Key concepts

Non-invasive Power Analysis
This technique measures the instantaneous electrical power consumption of a chip without physically altering its operation or requiring invasive probes. It captures the subtle fluctuations in current draw caused by different computational tasks running concurrently on the hardware, allowing researchers to infer which specific function is active.
Hierarchical Classification Architecture
The system uses a two-stage machine learning approach. The first stage specifically targets short operations like AES using a sliding window. If that fails, the second stage takes over with a multi-class classifier to distinguish between other tasks like ECC or CNN based on the full power trace characteristics.
Feature Extraction (Time and Frequency Domain)
The framework extracts 21 features from raw power traces. Time-domain metrics capture signal shape changes, while frequency-domain features, derived via FFT, highlight specific periodic switching patterns. These combined metrics are crucial for distinguishing the unique spectral signatures of different hardware accelerators.
Principal Component Analysis (PCA) Projection
PCA is used to reduce the complexity of 21 extracted features into a 2D space. This projection reveals how different workloads cluster together in terms of their power signature. It helps resolve ambiguities, such as separating AES from ECC, by showing their relative proximity in this lower-dimensional feature map.

Terminology used across episodes

This episode discusses

The paper

HAWC-Leak: Heterogeneous Architecture Workload Classification utilizing Side-Channel Leakage · Read on arXiv

Washington University in St. Louis · University of Utah

Modern heterogeneous computing platforms, which integrate general-purpose processors with specialized accelerators such as FPGAs, have become the standard for high-performance and embedded applications. However, the shared physical resources of these systems introduce critical security vulnerabilities, where power consumption patterns can leak sensitive runtime information. While side-channel analysis is well established for extracting cryptographic keys from isolated algorithms, existing methods struggle to profile complex, dynamic systems where diverse workloads coexist on shared hardware resources. This paper presents HAWC-Leak, a dynamic workload classification framework designed to identify specific hardware accelerator tasks in real time using non-invasive power analysis. Unlike prior studies focused on extracting cryptographic secrets from isolated implementations, our approach profiles heterogeneous runtime behavior using machine learning to predict active workloads from features extracted from power traces. We implement a complete hardware-in-the-loop measurement testbed using the ChipWhisperer CW305 Artix-7 FPGA platform to capture high-resolution power signatures of diverse workloads including Advanced Encryption Standard (AES-128), Elliptic Curve Cryptography (ECC P-256), Secure Hash Algorithm (SHA-256), Convolutional Neural Network (CNN), and a system idle state. We develop a Hierarchical Random Forest classifier that processes real-time power data to identify the executing application in controlled mixed-activity scenarios where multiple accelerators are instantiated but only one executes at a time. Experimental results demonstrate that this framework achieves 100% classification accuracy across all evaluated scenarios, confirming that power consumption patterns serve as a reliable and distinctive fingerprint for profiling heterogeneous runtime behavior.

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "HAWC-Leak: Heterogeneous Architecture Workload Classification utilizing Side-Channel Leakage".

Elias: Here's the gist of the paper. The gist: HAWC-Leak presents a dynamic workload classification framework that identifies specific hardware accelerator tasks in real time using non-invasive power analysis and achieves 100% classification accuracy across all evaluated scenarios.

Nadia: Modern computing systems are increasingly heterogeneous, integrating general-purpose CPUs with GPUs, FPGAs, and domain-specific AI accelerators to meet the escalating performance and energy-efficiency demands of emerging workloads.

Elias: These platforms introduce critical security vulnerabilities where power consumption patterns can leak sensitive runtime information.

Nadia: Existing methods struggle to profile complex, dynamic systems where diverse workloads coexist on shared hardware resources.

Title and authors: Nadia: Now that we know what the paper is about, let's look at how they actually set up this experiment in "HAWC-Leak: Heterogeneous Architecture Workload Classification utilizing Side-Channel Leakage". They built a complete measurement pipeline using the ChipWhisperer platform and a Xilinx Artix-seven FPGA <ref:2610.30400#pg2,platform and a Xilinx Artix-7 FPGA>.

Elias: They weren't just testing one thing; they emulated a realistic edge computing scenario by deploying several different IP cores on that FPGA, including AES-one hundred twenty-eight ECC P-two hundred fifty-six SHA-two hundred fifty-six and even a Convolutional Neural Network or CNN fourteen–sixteen <ref:2610.30400#pg2,including AES-128, ECC P-256, SHA-256, and>.

Priya: So they’re not just looking at a clean operation; they are generating controlled mixed activity power traces where multiple accelerators are on the same fabric but only one runs at a time.

Nadia: That's the key difficulty they address, because those conditions reflect real deployment environments where static and dynamic power components overlap and obscure individual signatures.

Elias: They used synchronous sampling with the ChipWhisperer to reduce timing jitter, which lets them capture high-frequency transient signals precisely without needing massive trace averaging <ref:2610.30400#pg3>.

Priya: And they measured the power on a specific rail, VCCINT, using a shunt resistor setup so they could get a direct look at the instantaneous current draw from the programmable logic fabric.

The paper's summary: Nadia: The paper summarizes their approach as creating this dynamic workload classification framework designed to identify specific hardware accelerator tasks in real time using non-invasive power analysis. It moves beyond traditional methods that rely on simple thresholding because they acknowledge the complex noise in mixed systems.

Elias: They specifically targeted the runtime behavior of these heterogeneous systems by leveraging machine learning to predict the active workload from features extracted from those power traces <ref:2610.30400#pg1>.

Priya: So, what does this actually mean for us? It’s that instead of just saying "there's activity," they are trying to pinpoint exactly *what* is causing that activity in a noisy setup.

Nadia: Right. They achieved this by using a two-stage Random Forest classifier to distinguish between short operations and longer workloads, which helps separate the signal from the noise.

Elias: The first stage is a binary AES detector that uses a sliding window approach—a window size of one hundred twenty-nine samples with a stride of thirty-two—to detect if anything looks like an AES operation versus just being idle <ref:2610.30400#pg2>.

Priya: And if the first stage comes back negative, meaning it’s not AES, the second stage takes over to discriminate between ECC, SHA-two hundred fifty-six Idle, and CNN using features from the whole trace duration <ref:2610.30400#pg2>.

The paper's improvements: Nadia: One major improvement they suggest is this hierarchical model structure. It tackles the ambiguity that arises when different workloads have overlapping power envelopes, which traditional classifiers struggle with.

Elias: They found that for Stage one spectral rolloff and spectral centroid were the most important features for distinguishing AES from other things because those features capture the high-frequency switching characteristics <ref:2610.30400#pg2>.

Priya: For Stage two they shifted to time-domain metrics like peak-to-peak ratio and crest factor to discriminate between ECC, SHA-two hundred fifty-six and CNN based on the amplitude profile of those traces <ref:2610.30400#pg2>.

Nadia: They also used Principal Component Analysis or PCA on twenty-one extracted feature vectors to map the power signatures into a two-dimensional latent space. This mapping shows that AES and ECC clusters are physically closer together in this space, which justifies why they needed that hierarchical setup <ref:2610.30400#pg2>.

Elias: They also added a robustness improvement by planning to incorporate unsupervised learning methods like Invariant Information Clustering or IIC-SCA to make the system less dependent on perfectly labeled training sets.

Conclusion: Nadia: So, to wrap up this paper on "HAWC-Leak: Heterogeneous Architecture Workload Classification utilizing Side-Channel Leakage", the main point is that distinct accelerator workloads do have identifiable power signatures even when there's background leakage from idle modules.

Elias: They’ve established a strong baseline for scalable workload fingerprinting using non-invasive power side-channel analysis, confirming that power consumption acts like a highfidelity barcode for hardware execution <ref:2610.30400#pg2>.

Priya: It confirms that this method works even when multiple accelerators are running on the same fabric, as long as only one executes at a time, which is a critical real-world condition.

Nadia: The framework supports identifying workloads at least for single active workloads in this test, and it sets up a solid foundation for future work in profiling complex systems.

Elias: We should keep an eye on how they incorporate those unsupervised learning methods they mentioned to see if it makes the system more adaptable to different operational profiles.

Priya: It shows that even with noise, you can build a way to reliably profile what's happening at the hardware level without needing perfect knowledge of every possible scenario.

More episodes

← Home