CIRISOssicle

GPU workload detection via timing-based strain gauge.

Software-only detection of unauthorized workloads (crypto mining, resource hijacking) - no external hardware required.

┌────────────────────────────────────────────────────────────────┐
│                      CIRISOssicle                              │
│                                                                │
│   Kernel ──► Timing ──┬──► WORKLOAD [VALIDATED]                │
│              4000 Hz  │    Mean shift >20%, latency 2.5 ms     │
│                       │                                        │
│                       ├──► EMI/VDF/THERMAL [THEORIZED]         │
│                       │    Spectrum analysis, not yet proven   │
│                       │                                        │
│                       └──► TRNG (7.99 bits/byte)               │
│                                                                │
│   Workload detection: O1-O7 Validated January 2026             │
└────────────────────────────────────────────────────────────────┘

License: Business Source License 1.1 (BSL 1.1) Licensor: CIRIS L3C (Eric Moore) Change Date: January 1, 2030 Change License: AGPL 3.0

Free for individuals, DIY, academics, nonprofits, and orgs <$1M revenue.

What's Validated vs Theorized

Capability	Status	Evidence
Workload Detection	VALIDATED	O1-O7 experiments, 100% TP, 0% FP
EMI Detection	Theorized	Spectrum analysis implemented, not tested
VDF Detection	Theorized	Band 20-100 Hz hypothesized
Thermal Detection	Theorized	ACF drift observed, not characterized

Workload Detection Results (O1-O7 Validated)

Detection uses mean shift at 4000 Hz sample rate:

Workload	Mean Shift	CV	Detection
1%	+192%	-	100% TP
30%	+188%	13.6%	100% TP
50%	+248%	13.6%	100% TP
70%	+317%	3.4%	100% TP
90%	+380%	-	100% TP

Key Specs:

Metric	Value	Source
Sample rate	4000 Hz	O2
Detection latency	2.5 ms	O4
Detection floor	1% workload	O5
Mean shift at 50%	+248%	O5
CV at 70% load	3.4%	O7
Avoid zone	1900-2100 Hz	O2b

Quick Start

pip install cupy-cuda12x numpy scipy

# Production strain gauge
python src/strain_gauge.py

# Run validation experiments
python experiments/o1_replicate_b1e.py
python experiments/o3_o5_4khz_validation.py
python experiments/o4_o7_latency_cv.py

Usage

from src.strain_gauge import StrainGauge

# Workload detection
gauge = StrainGauge()
gauge.calibrate(duration=10.0)  # Includes warm-up
reading = gauge.read()
if reading.detected:
    print(f"WORKLOAD DETECTED: mean_shift={reading.mean_shift_pct:.1f}%")

# Check system health (ACF should be ~0.45)
print(f"State: {reading.system_state}, ACF: {reading.acf:.2f}")

# TRNG
trng_result = gauge.generate_trng(16)
print(f"Random: {trng_result.bytes.hex()}")

How It Works

Detection is based on mean shift monitoring at 4000 Hz:

1. Run minimal GPU kernel, measure timing (nanoseconds)
2. Sample at 4000 Hz for optimal signal-to-noise
3. Detect workloads via mean_shift > 20%
4. GPU contention causes timing to triple (~7μs → ~21μs)
5. Detection latency: 2.5 ms (10ms window)

Key insights:

• 4000 Hz has lowest variance (±14% vs ±41% at 1000 Hz)
• Avoid 1900-2100 Hz (interference dip at ~2050 Hz)
• Detection floor is 1% workload (+192% mean shift)

Validation Summary (O1-O7)

Exp	Test	Result	Status
O1	Mean shift at 1790 Hz	+153%	PASS
O2	Optimal sample rate	4000 Hz	PASS
O2b	Interference dip	~2050 Hz	DOCUMENTED
O3	Workload band at 4000 Hz	7.1%	PASS
O4	Detection latency	2.5 ms	PASS
O5	Detection floor	1%	PASS
O7	CV at 70% load	3.4%	PASS

Distribution Finding

Z-scores are Student-t distributed (NOT Gaussian):

• Kurtosis κ = 230
• Degrees of freedom ≈ 1.3
• Individual z-scores unreliable, use mean shift

Prior Art

Work	Year	Relevance
US9459834B2	2011	GPU TRNG via timing (expired 2024)
GPUs and chaos: TRNG	2015	Chaos + timing for RNG
ShadowScope	2025	GPU monitoring via side channels
GPU Cryptojacking Detection	2023	ML-based detection

What's different here: Lightweight implementation (256 bytes), no CUPTI/nvidia-smi dependency, 2.5ms detection latency, 1% detection floor.

Limitations

• CV is higher at lower intensities (13.6% at 30-50% vs 3.4% at 70%)
• Avoid 1900-2100 Hz sample rates (GPU-specific interference)
• Validated on RTX 4090 only (other GPUs may differ)
• CPU-only attacks not detected
• Timing-aware evasion possible for sophisticated attackers

Related Projects

Project	Description
RATCHET	GPU Lorenz oscillator research (same findings)
CIRISArray	Multi-sensor detection

Requirements

• NVIDIA GPU with CUDA
• CuPy (pip install cupy-cuda12x)
• NumPy, SciPy

Why "Ossicle"?

The ossicles are tiny bones in the middle ear that amplify vibrations. The name reflects the original goal of maximum sensitivity from minimum size.

Citation

@software{cirisossicle2026,
  title = {CIRISOssicle: GPU Workload Detection via Mean Shift},
  author = {CIRIS L3C},
  year = {2026},
  license = {BSL-1.1},
  note = {O1-O7 validated: 4000 Hz, 2.5ms latency, 1% floor}
}