masterpieceexternal/gpt-oss-20b-moe-expert-power-traces-320k

GPT-OSS-20B MoE Expert Power Traces (320k, ChipWhisperer)

This dataset contains analog power traces captured with a ChipWhisperer Husky while running forced single-expert MoE computations derived from openai/gpt-oss-20b on an NVIDIA H100.

What is recorded

Each trace corresponds to one capture trial where:

1. A fixed expert id is selected (expert_00 ... expert_31).
2. A random hidden-state tensor is generated once per trial.
3. The selected expert computation is executed repeatedly inside one capture window (expert_iters=12).
4. ChipWhisperer records a ~10 ms analog trace from the power sensing setup.

Important: this is not a full unmodified model forward pass. It is a controlled harness for expert-identification side-channel experiments.

Dataset layout

• capture_meta.json: capture configuration and metadata
• traces/expert_XX/trial_YYYYYY.npy: raw captured trace for a class/trial

Class count: 32 experts (expert_00..expert_31)

Samples per class: 10,000

Total traces: 320,000

Trace format

• File type: NumPy .npy
• Array dtype: floating-point (captured analog samples)
• Typical duration: ~10 ms per trace
• Captures include repeated expert activity inside one window (12 repetitions)

Baseline training recipe used in experiments

A common preprocessing/training setup used with this dataset:

• Baseline normalization from early-trace samples
• Resample trace to fixed feature length (e.g., 16,384)
• Add first-difference channel (dx)
• Train 1D CNN for 32-way expert classification

Known caveats

• No pre-trigger idle segment in this capture run.
• Early samples may include launch/ramp transients depending on timing.
• Repetition within a trace means each sample is a composite of multiple expert invocations.
• GPU state drift (clock/thermal/cache) can introduce non-stationarity.

Intended use

• Side-channel feasibility studies for MoE expert identification
• Feature engineering and leakage-localization experiments
• Benchmarking robust time-series classifiers under drift/jitter

Ethical and security note

This dataset is released for defensive research and measurement methodology work. Do not use it to target systems without authorization.

Included collection script

• scripts/train_expert_classifier_multiclass.py: script used to run capture/training workflows; this dataset was captured with its multiclass expert-trace capture path and corresponding arguments.