Card Anomaly Detection Eval

card-anomaly-detection-eval

Benchmarking the Energy Cost of Assurance in Neuromorphic Edge Robotics — Kaczmarek (2026) (arXiv:2603.13880, 2026)

What this evaluates

Evaluates the robustness and energy efficiency of a neuromorphic spiking neural network for real-time anomaly detection on lunar rover sensor telemetry. It specifically probes the model's ability to maintain classification accuracy under gradient-based and temporal adversarial attacks while measuring hardware-level power consumption.

Datasets

• Cislunar Anomaly and Risk Dataset (CARD) — total ?; splits: test (-1)

Metrics

• Adversarial Success Rate (ASR) (primary) — range: percent
  • Percentage of attack attempts that successfully cause the model to misclassify an anomaly as a nominal state.
• Energy per Inference — range: other
  • Integrated instantaneous power draw over the inference window, divided by the number of samples processed. Dynamic energy subtracts an idle baseline.
• Latency — range: other
  • Wall-clock time required for the hardware to process a single input window and output a classification event, excluding PCIe transfer overhead.

Input / output format

Input: Streaming multi-modal sensor telemetry (IMU data and visual feeds) processed as input windows for real-time classification.

Output: Binary classification event indicating system health (nominal vs. anomaly).

Scoring recipe

def compute_asr(predictions, gold_labels, attack_mask):
    successful = sum(1 for p, g, a in zip(predictions, gold_labels, attack_mask) if a and p != g and g == 'anomaly')
    return (successful / sum(attack_mask)) * 100

def compute_energy(power_samples, sample_rate, num_samples, idle_baseline):
    total_energy = sum(power_samples) / sample_rate
    return (total_energy / num_samples) - idle_baseline

Common pitfalls

• Energy and latency measurements must exclude PCIe transfer overhead to isolate the neuromorphic processor's performance.
• Adversarial Success Rate (ASR) only counts misclassifications where an anomaly is incorrectly labeled as nominal, not all errors.
• All metrics are averaged over 10 independent runs and measured with batch size 1 to simulate real-time edge processing.

Evidence (verbatim from paper)

We define the Adversarial Success Rate (ASR) as the percentage of attack attempts that successfully cause the model to misclassify an anomaly as a nominal state. The total energy per inference is calculated by integrating the instantaneous power draw over the duration of the inference window and dividing by the number of samples processed.

Citation

@misc{kaczmarek2026benchmarking,
  title={Benchmarking the Energy Cost of Assurance in Neuromorphic Edge Robotics},
  author={Kaczmarek (2026)},
  year={2026},
  note={arXiv:2603.13880}
}

• arXiv: 2603.13880