Q&A: Phase 38.2 DifferentialPrivacyEngine — Noise Calibration & Budget Management

[web3guru888]

Q&A: DifferentialPrivacyEngine (Phase 38.2)

Configuration Questions

Q1: How to set the noise multiplier σ?

• Use get_noise_multiplier() — specify target (ε,δ), sampling rate, and number of steps
• Typical values: σ ∈ [0.5, 2.0] for ε ∈ [1, 10]
• Higher σ → stronger privacy, lower accuracy
• σ = 1.1 with C = 1.0 is a good starting point (Abadi et al.)

Q2: What clipping bound C should I use?

• Too small: lose signal, slow convergence
• Too large: weak privacy (noise scales with C)
• Adaptive clipping: set C to the median per-sample gradient norm
• Start with C = 1.0, tune based on gradient norm distribution

Q3: Gaussian vs Laplace mechanism?

• Gaussian: better for high-dimensional outputs (L2 sensitivity), requires δ > 0
• Laplace: pure ε-DP (δ = 0), better for low-dimensional queries (L1 sensitivity)
• For deep learning: always Gaussian (high-dimensional gradient vectors)

Q4: How many training rounds can I afford with privacy budget ε = 8?

• Depends on σ, sampling rate q, and accounting method
• Example: σ=1.1, q=0.01, δ=1e-5 → ~500 rounds for ε=8 (RDP accountant)
• Moments accountant gives tighter bounds than basic composition by 10-100x

Q5: When does the privacy budget run out?

• check_budget() returns False when next operation would exceed target ε
• Configure estimated_rounds_left for proactive planning
• Strategy: allocate 80% of budget to training, 20% to evaluation

See issue #773 for full specification and acceptance criteria.