Homomorphic Svm Eval

---
name: homomorphic-svm-eval
description: Evaluates the accuracy, latency, and energy efficiency of a homomorphic inference accelerator performing SVM classification on encrypted data under intermittent power constraints. Use when the user wants to benchmark on MNIST, Human Activity Recognition, ADULT, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2112.08943
  bibtex_key: resch2021homomorphic
  confidence: high
---

# homomorphic-svm-eval

> Towards Homomorphic Inference Beyond the Edge — Resch et al. (2021) (arXiv:2112.08943, 2021)

## What this evaluates

Evaluates the accuracy, latency, and energy efficiency of a homomorphic inference accelerator performing SVM classification on encrypted data under intermittent power constraints.

## Datasets

- **MNIST** — total ?; splits: test (-1)
- **Human Activity Recognition** — total ?; splits: test (-1)
- **ADULT** — total ?; splits: test (-1)

## Metrics

- `accuracy` **(primary)** — range: percent
  - Percentage of correctly classified samples after decrypting the homomorphic SVM output, compared against ground truth labels.
- `latency` — range: other
  - Total time in seconds to complete one inference, including data transmission, encoding, homomorphic computation, and decryption.
- `energy` — range: other
  - Total microjoules consumed per inference, measured via hardware simulator extrapolated from NVSIM and theoretical power models.

## Input / output format

**Input**: Fixed-length ciphertexts of 4096 elements containing normalized 3-bit integer inputs (values 0-7), padded with dummy data to fill the ciphertext. Each input element occupies a separate ciphertext slot.

**Output**: Encrypted SVM prediction, which is decrypted to a class label. Accuracy is computed by comparing decrypted labels to ground truth.

## Scoring recipe

```python
predictions = [decrypt(rodent_inference(enc_x)) for enc_x in encrypted_inputs]
accuracy = sum(p == y for p, y in zip(predictions, labels)) / len(labels)
# Latency and energy are measured directly from the in-house hardware simulator rather than computed from predictions.
```

## Common pitfalls

- Ciphertexts are always padded to 4096 elements regardless of actual input dimension, which inflates computation and energy metrics compared to the raw input size.
- Inputs are quantized to 3-bit integers (0-7) to prevent arithmetic overflow, which inherently limits maximum achievable accuracy compared to full-precision models.
- Latency and energy figures are derived from a custom simulator extrapolated from NVSIM and theoretical power models, not from physical silicon measurements.

## Evidence (verbatim from paper)

> Despite these limitations, we were still able to achieve reasonable accuracy relative to full-precision SVMs provided by libSVM [[13]], which we accessed through R [[73]] with the ‘e1071’ package [[62]]. The SVM parameters and accuracies are listed in Table I.

## Citation

```bibtex
@misc{resch2021homomorphic,
  title={Towards Homomorphic Inference Beyond the Edge},
  author={Resch et al. (2021)},
  year={2021},
  note={arXiv:2112.08943}
}
```

- arXiv: 2112.08943