Google is making private AI practical with homomorphic encryption
Google’s push to use fully homomorphic encryption (FHE) for “private AI” aims to let cloud models run on user-encrypted data so providers never see the plaintext, potentially unlocking regulated use cases in healthcare, finance, and other sensitive domains. Commenters acknowledge the cryptographic soundness and niche value of FHE but stress its severe performance overhead — often orders of magnitude slower — and note that running models locally is still the most straightforward way to get privacy. Many are also skeptical that an ad-driven company like Google will deploy such technology in ways that truly prioritize user privacy over data collection and monetization.
Overall sentiment
- Mixed to negative on Google as steward; more positive on the underlying cryptography.
- Many see this as technically impressive but commercially niche and potentially mis-framed as “private AI” while keeping control in the cloud.
Trust, motives, and privacy framing
- Strong distrust of Google as an advertising company; fear FHE will be used to justify more pervasive data use (“we never see your data, just the signals”).
- Some argue the project is open source, so the tech can benefit others and need not rely on trusting Google.
- Several note that encryption only solves confidentiality; you still lose availability/control if your account is blocked or service is shut down.
What FHE actually provides (and does not)
- FHE lets servers compute on ciphertext without seeing plaintext, assuming standard hardness assumptions (e.g., LWE/RLWE).
- It guarantees secrecy of inputs/outputs, but not that the server ran the intended computation. Verifiable/attested compute is a separate problem.
- Clarifications that proper FHE can still satisfy indistinguishability-from-noise (IND-CPA); multiple commenters correct misunderstandings here.
Performance and practicality
- Overheads cited from ~10×–100× (optimistic for tuned ML) up to 10³–10⁶× in many real setups; benchmarked latencies from seconds to minutes per inference or operation.
- Sorting, branching, and division are particularly slow; linear algebra and simple add/mul workloads are relatively friendlier.
- Ongoing work on GPU and ASIC acceleration is mentioned, but many still see it as far from viable for large LLMs.
Use cases vs. local compute
- Suggested applications: biometric auth, password leak checking, medical and DNA queries, financial fraud detection, inter-bank transfers, private ad targeting, “have I been pwned”-style checks.
- Some argue most of these are better solved by local compute, TEEs, or legal/contractual controls, especially given cost.
- Others note FHE can unlock regulated workloads (healthcare, finance) where local or on-prem compute is not always feasible or where multi-party data aggregation is needed.
Alternative models and ecosystem
- Comparisons with secure enclaves/TEEs: cheaper and faster but rely on hardware vendor and operator trust, and are frequently broken via side channels.
- Several see this as necessary “programmable cryptography” research, even if near-term deployment is narrow. Others dismiss it as Google “research theater” to impress AI-focused executives.