A comparative analysis of noise schemes for Fully Homomorphic Encryption on resource-constrained edge devices — covering performance, power, and memory trade-offs.
Fully Homomorphic Encryption (FHE) enables privacy-preserving inference on edge devices by allowing computation on encrypted data. However, FHE's practical deployment is bottlenecked by the noise distribution chosen during encryption — directly impacting latency, power draw, and memory footprint.
This white paper compares six FHE noise distributions evaluated against the constraints of resource-limited edge hardware, identifying Centered Binomial Noise as the optimal default and Sparse Noise as the best option for ultra-constrained environments.
FHE introduces computational overhead that scales dramatically with noise complexity. On edge devices operating under strict power (< 250 mW), memory (< 12 MB), and latency (< 200 ms) budgets, the choice of noise distribution becomes a first-order architectural decision.
Most FHE literature focuses on cryptographic security proofs rather than hardware-aware performance profiling — leaving edge system architects without actionable guidance for noise scheme selection.
| Noise Type | Category | Key Characteristic |
|---|---|---|
| Gaussian | Continuous | Standard FHE baseline; high precision, high cost |
| Centered Binomial | Discrete, HW-Friendly | Best overall balance for edge hardware |
| Ternary {-1, 0, +1} | Discrete | Simple structure; moderate across all metrics |
| Uniform | Continuous | Easy to sample; resource-intensive in practice |
| Sparse (Mostly Zeros) | Discrete | Lowest memory; ideal for extreme constraints |
| Discrete Laplace | Heavy-Tailed | Privacy-focused; moderate–slow performance |
| Noise Type | Latency | Rating | Power | Rating | Memory | Rating |
|---|---|---|---|---|---|---|
| Gaussian | 120–180 ms | Slow | 180–250 mW | High | 8–12 MB | High |
| Centered Binomial | 40–80 ms | Fast | 60–100 mW | Low | 2–4 MB | Low |
| Ternary | 70–120 ms | Moderate | 90–150 mW | Moderate | 3–6 MB | Moderate |
| Uniform | 100–160 ms | Slow | 150–220 mW | High | 6–10 MB | High |
| Sparse | 50–90 ms | Fast | 70–120 mW | Low | 1–3 MB | Very Low |
| Discrete Laplace | 90–140 ms | Mod–Slow | 120–180 mW | Mod–High | 4–8 MB | Moderate |
All metrics normalized to Gaussian baseline (1.00). Lower values indicate better performance.
Weighted composite: 40% Latency · 30% Power · 30% Memory. Lower score = better.
| Rank | Noise Type | Score |
|---|---|---|
| 1 | Centered Binomial | 0.28 |
| 2 | Sparse | 0.30 |
| 3 | Ternary | 0.47 |
| 4 | Discrete Laplace | 0.62 |
| 5 | Uniform | 0.83 |
| 6 | Gaussian | 1.00 |
Best overall balance for edge AI FHE — significantly lower latency, power, and memory with strong hardware support. Default recommendation for most deployments.
Achieves the lowest memory (1–3 MB) and competitive power. Ideal for highly constrained devices when the FHE scheme supports it.
Moderate trade-offs across all metrics. Viable alternatives when specific scheme compatibility or differential privacy integration is required.
Most resource-intensive distributions. Less suitable for tight edge environments despite prevalence in theoretical FHE literature.
| Factor | Guidance |
|---|---|
| FHE Scheme | Verify noise support in target library (SEAL, OpenFHE, Lattigo) |
| Hardware | ARM Cortex-M / RISC-V favor Centered Binomial; FPGA can exploit Sparse structure |
| Security Level | All distributions meet 128-bit security at recommended parameters |
| Maturity | Centered Binomial has widest library support; Sparse may need custom kernels |
| Power Budget | < 100 mW → Centered Binomial or Sparse · > 150 mW → Ternary viable |
Efficient noise selection is critical for enabling practical, secure, and sustainable Edge AI with FHE. Centered Binomial Noise emerges as the clear winner for general-purpose edge FHE — offering a 4× improvement over Gaussian baselines in latency, power, and memory simultaneously.
For ultra-constrained endpoints, Sparse Noise provides an even leaner alternative. Qoresic's edge AI platform incorporates these findings into its SRAM-optimized inference engine, enabling privacy-preserving intelligence on devices operating under 1 W.